The solutions to many computer vision problems involve the estimation of some visual properties of an image or video, represented as either discrete or continuous variables. For example scene classification aims to estimate the value of a discrete variable indicating which scene category an image belongs to; for object detection the task is to estimate a binary variable corresponding the presence/absence of the object of interest and a set of variables indicating its whereabouts in the image plane (e.g. four variables if the whereabouts are represented as bounding boxes).Most of these visual properties are objective; that is, there is no or little ambiguity in their true values to a human annotator.
In comparison, the problem of estimating subjective visual properties is much less studied. This class of computer vision problems nevertheless encompass a variety of important applications. For example: estimating attractiveness  from faces would interest social media or online dating websites; and estimating properties of consumer goods such as shininess of shoes  improves customer experiences on online shopping websites. Recently, the problem of automatically predicting if people would find an image or video interesting has started to receive increasing attention [3, 4, 5]. Interestingness prediction has a number of real-world applications. In particular, since the number of images and videos uploaded to the Internet is growing explosively, people are increasingly relying on image/video recommendation tools to select which ones to view. Given a query, ranking the retrieved data with relevance to the query based on the predicted interestingness would improve user satisfaction. Similarly user stickiness can be increased if a media-sharing website such as YouTube can recommend videos that are both relevant and interesting. Other applications such as web advertising and video summarisation can also benefit. Subjective visual properties such as the above-mentioned ones are useful on their own. But they can also be used as an intermediate representation for other tasks such as visual recognition, e.g., different people can be recognised by how pale their skin complexions are and how chubby their faces are . When used as a semantically meaningful representation, these subjective visual properties often are referred to as relative attributes [2, 6, 7].
Learning a model for subjective visual property (SVP) prediction is challenging primarily due to the difficulties in obtaining annotated training data. Specifically, since most SVPs can be represented as continuous variables (e.g. an interestingness/aesthetics/shininess score with a value range of 0 to 1 with 1 being most interesting/aesthetically appealing/shinning), SVP prediction can be cast as a regression problem – the low-level feature values are regressed to the SVP values given a set of training data annotated with their true SVP values. However, since by definition these properties are subjective, different human annotators often struggle to give an absolute value and as a result the annotations of different people on the same instance can vary hugely. For example, on a scale of 1 to 10, different people will have very different ideas on what a scale 5 means for an image, especially without any common reference point. On the other hand, it is noted that humans can in general more accurately rank a pair of data points in terms of their visual properties [8, 9] , e.g. it is easier to judge which of two images is more interesting relatively than giving an absolute interestingness score to each of them. Most existing studies [2, 1, 9] on SVP prediction thus take a learning to rank approach , where annotators give comparative labels about pairs of images/videos and the learned model is a ranking function that predicts the SVP value as a ranking score.
To annotate these pairwise comparisons, crowdsourcing tools such as Amazon Mechanic Turk (AMT) are resorted to, which allow a large number of annotators to collaborate at very low cost. Data annotation based on crowdsourcing is increasingly popular [6, 2, 4, 5] recently for annotating large-scale datasets. However, this brings about two new problems: (1) Outliers – The crowd is not all trustworthy: it is well known that crowdsourced data are greatly affected by noise and outliers [11, 12, 13] which can be caused by a number of factors. Some workers may be lazy or malicious , providing random or wrong annotations either carelessly or intentionally; some other outliers are unintentional human errors caused by the ambiguous nature of the data, thus are unavoidable regardless how good the attitudes of the workers are. For example, the pairwise ranking for Figure 1(a) depends on the cultural/psychological background of the annotator – whether s/he is more familiar/prefers the story of Monkey King or Cookie Monster111This is also known as Halo Effect in Psychology.. When we learn the model from labels collected from many people, we essentially aim to learn the consensus, i.e. what most people would agree on. Therefore, if most of the annotators growing up watching Sesame Street thus consciously or subconsciously consider the Cookie Monster to be more interesting than the Monkey King, their pairwise labels/votes would represent the consensus. In contrast, one annotator who is familiar with the stories in Journey to the West may choose the opposite; his/her label is thus an outlier under the consensus. (2) Sparsity – the number of pairwise comparisons required is much bigger than the number of data points because instances define a pairwise space. Consequently, even with crowdsourcing tools, the annotation remains be sparse, i.e. not all pairs are compared and each pair is only compared a few times.
To deal with the outlier problem in crowdsourced data, existing studies take a majority voting strategy [6, 2, 4, 15, 16, 17, 18]. That is, a large budget of times the number of actual annotated pairs required is allocated to obtain multiple annotations for each pair. These annotations are then averaged over so as to eliminate label noise. However, the effectiveness of the majority voting strategy is often limited by the sparsity problem – it is typically infeasible to have many annotators for each pair. Furthermore, there is no guarantee that outliers, particularly those caused by unintentional human errors can be dealt with effectively. This is because majority voting is a local consistency detection based strategy – when there are contradictory/inconsistent pairwise rankings for a given pair, the pairwise rankings receiving minority votes are eliminated as outliers. However, it has been found that when pairwise local rankings are integrated into a global ranking, it is possible to detect outliers that can cause global inconsistency and yet are locally consistent, i.e. supported by majority votes . Critically, outliers that cause global inconsistency have more significant detrimental effects on learning a ranking function for SVP prediction and thus should be the main focus of an outlier detection method.
In this paper we propose a novel approach to subjective visual property prediction from sparse and noisy pairwise comparison labels collected using crowdsourcing tools. Different from existing approaches which first remove outliers by majority voting, followed by regression  or learning to rank , we formulate a unified robust learning to rank (URLR) framework to solve jointly both the outlier detection and learning to rank problems. Critically, instead of detecting outliers locally and independently at each pair by majority voting, our outlier detection method operates globally, integrating all local pairwise comparisons together to minimise a cost that corresponds to global inconsistency of ranking order. This enables us to identify those outliers that receive majority votes but cause large global ranking inconsistency and thus should be removed. Furthermore, as a global method that aggregates comparisons across different pairs, our method can operate with as few as one comparison per pair, making our method much more robust against the data sparsity problem compared to the conventional majority voting approach that aggregates comparisons for each pair in isolation. More specifically, the proposed model generalises a partially penalised LASSO optimisation or Huber-LASSO formulation [20, 21, 22] from a robust statistical ranking formulation to a robust learning to rank model, making it suitable for SVP prediction given unseen images/videos. We also formulate a regularisation path based solution to solve this new formulation efficiently. Extensive experiments are carried out on benchmark datasets including two image and video interestingness datasets [4, 5] and two relative attribute datasets . The results demonstrate that our method significantly outperforms the state-of-the-art alternatives.
2 Related work
Subjective visual properties Subjective visual property prediction covers a large variety of computer vision problems; it is thus beyond the scope of this paper to present an exhaustive review here. Instead we focus mainly on the image/video interestingness prediction problem which share many characteristics with other SVP prediction problem such as image quality , memorability , and aesthetics  prediction.
Predicting image and video interestingness Early efforts on image interestingness prediction focus on different aspects than interestingness as such, including memorability  and aesthetics . These SVPs are related to interestingness but different. For instance, it is found that memorability can have a low correlation with interestingness - people often remember things that they find uninteresting . The work of Gygli et al  is the first systematic study of image interestingness. It shows that three cues contribute the most to interestingness: aesthetics, unusualness/novelty and general preferences, the last of which refers to the fact that people in general find certain types of scenes more interesting than others, for example outdoor-natural vs. indoor-manmade. Different features are then designed to represent these cues as input to a prediction model. In comparison, video interestingness has received much less attention, perhaps because it is even harder to understand its meaning and contributing cues. Liu et al.  focus on key frames so essentially treats it as an image interestingness problem, whilst  is the first work that proposes benchmark video interestingness datasets and evaluates different features for video interestingness prediction.
Most earlier works cast the aesthetics or interestingness prediction problem as a regression problem [23, 3, 24, 25]. However, as discussed before, obtaining an absolute value of interestingness for each data point is too subjective and affected too much by unknown personal preference/social background to be reliable. Therefore the most recent two studies on image  and video  interestingness all collect pairwise comparison data by crowdsourcing. Both use majority voting to remove outliers first. After that the prediction models differ –  converts pairwise comparisons into an absolute interestingness values and use a regression model, whilst  employs rankSVM  to learn a ranking function, with the estimated ranking score of an unseen video used as the interestingness prediction. We compare with both approaches in our experiments and demonstrate that our unified robust learning to rank approach is superior as we can remove outliers more effectively – even if they correspond to comparisons receiving majority votes, thanks to its global formulation.
instances and classes at an intermediate level of representation. Attributes are then used for various tasks including N-shot and zero-shot transfer learning. Most previous studies consider binary attributes[26, 27]. Relative attributes  were recently proposed to learn a ranking function to predict relative semantic strength of visual attributes. Instead of the original class-level attribute comparisons in , this paper focuses on instance-level comparisons due to the huge intra-class variations in real-world problems. With instance-level pairwise comparisons, relative attributes have been used for interactive image search , and semi-supervised 29, 30] of visual categories. However, no previous work addresses the problem of annotation outliers except 
, which adopts the heuristic majority voting strategy.
Learning from noisy paired crowdsourced data Many
large-scale computer vision problems rely on human intelligence tasks
(HIT) using crowdsourcing services, e.g. AMT (Amazon Mechanical Turk)
to collect annotations. Many studies [14, 31, 32, 13]
highlight the necessity of validating the random or malicious labels/workers
and give filtering heuristics for data cleaning. However, these are primarily based on majority voting
which requires a costly volume of redundant annotations, and has
no theoretical guarantee of solving the outlier and sparsity problems.
As a local (per-pair) filtering method, majority voting does not respect global ordering and even risks introducing additional
inconsistency due to the well-known Condorcet’s paradox in social choice and voting theory .
Active learning [34, 29, 30] is an another way to circumvent the pairwise labelling space. It actively poses specific requests to annotators and learns from their feedback, rather than the ‘general’ pairwise comparisons discussed in this work.
Besides paired crowdsourced data, majority voting is more widely used in crowdsourcing where multiple annotators directly label instances, which attracted lots of attention in the machine learning community
Besides paired crowdsourced data, majority voting is more widely used in crowdsourcing where multiple annotators directly label instances, which attracted lots of attention in the machine learning community[16, 17, 18, 15]. In contrast, our work focuses on pairwise comparisons which are relatively easier for annotators in evaluating the subjective visual properties  .
Statistical ranking and learning to rank Statistical ranking has been widely studied in statistics and computer science [35, 36, 8, 37]. However, statistical ranking only concerns the ranking of the observed/training data, but not learning to predict unseen data by learning ranking functions. To learn ranking functions for applications such as interestingness prediction, a feature representation of the data points must be used as model input in addition to the local ranking orders. This is addressed in learning to rank which is widely studied in machine learning [38, 39, 40]. However, existing learning to rank works do not explicitly model and remove outliers for robust learning: a critical issue for learning from crowdsourced data in practice. In this work, for the first time, we study the problem of robust learning to rank given extremely noisy and sparse crowdsourced pairwise labels. We show both theoretically and experimentally that by solving both the outlier detection and ranking prediction problems jointly, we achieve better outlier detection than existing statistical ranking methods and better ranking prediction than existing learning to rank method such as RankSVM without outlier detection.
Our contributions are threefold: (1) We propose a novel robust learning to rank method for subjective visual property prediction using noisy and sparse pairwise comparison/ranking labels as training data. (2) For the first time, the problems of detecting outliers and estimating linear ranking models are solved jointly in a unified framework. (3) We demonstrate both theoretically and experimentally that our method is superior to existing majority voting based methods as well as statistical ranking based methods. An earlier and preliminary version of this work is presented in  which focused only on the image/video interestingness prediction problem.
3 Unified Robust Learning to Rank
3.1 Problem definition
We aim to learn a subjective visual property (SVP) prediction model from a set of sparse and noisy pairwise comparison labels, each comparison corresponding to a local ranking between a pair of images or videos. Suppose our training set has data points/instances represented by a feature matrix , where is a
-dimensional column low-level feature vector representing instance. The pairwise comparison labels (annotations collected using crowdsourcing tools) can be naturally represented as a directed comparison graph with a node set corresponding to the instances and an edge set corresponding to the pairwise comparisons.
The pairwise comparison labels can be provided by multiple annotators. They are dichotomously saved: Suppose annotator gives a pairwise comparison for instance and (). If considers that the SVP of instance is stronger/more than that of , we save and set . If the opposite is the case, we save and set . All the pairwise comparisons between instances and are then aggregated over all annotators who have cast a vote on this pair; the results are represented as which is the total number of votes on over for a specific SVP, where indicates the Iverson’s bracket notation, and which is defined similarly. This gives an edge weight vector where is the number of edges. Now the edge set can be represented as and is the weight for the edge . In other words, an edge : exists if . The topology of the graph is denoted by a flag indicator vector where each indicator indicates that there is an edge between instances to regardless how many votes it carries. Note that all the elements in have the value , and their index gives the corresponding nodes in the graph.
Given the training data consisting of the feature matrix and the annotation graph , there are two tasks:
Detecting and removing the outliers in the edge set of . To this end, we introduce a set of unknown variables where each variable indicates whether the edge is an outlier. The outlier detection problem thus becomes the problem of estimating .
Estimating a prediction function for SVP. In this work a linear model is considered due to its low computational complexity, that is, given the low-level feature of a test instance we use a linear function to predict its SVP, where is the coefficient weight vector of the low-level feature . Note that all formulations can be easily updated to use a non-linear function.
So far in the introduced notations three vectors share indices: the flag indicator vector , the outlier variable vector and the edge weight vector . For notation convenience, from now on we use , and to replace , and respectively. As in most graph based model formulations, we define as the incident matrix of the directed graph , where if the edge enters/leaves vertex .
Note that in an ideal case, one hopes that the votes received on each pair are unanimous, e.g. and ; but often there are disagreements, i.e. we have both and . Assuming both cannot be true simultaneously, one of them will be an outlier. In this case, one is the majority and the other minority which will be pruned by the majority voting method. This is why majority voting is a local outlier detection method and requires as many votes per pair as possible to be effective (the wisdom of a crowd).
3.2 Framework formulation
In contrast to majority voting, we propose to prune outliers globally and jointly with learning the SVP prediction function. To this end, the outlier variables for outlier detection and the coefficient weight vector for SVP prediction are estimated in a unified framework. Specifically, for each edge , its corresponding flag indicator is modelled as
is the Gaussian noise with zero mean and a variance, and the outlier variable is assumed to have a higher magnitude than . For an edge , if is not an outlier, we expect should be approximately equal to , therefore we have . On the contrary, when the prediction of differs greatly from , we can explain as an outlier and compensate for the discrepancy between the prediction and the annotation with a nonzero value of . The only prior knowledge we have on is that it is a sparse variable, i.e. in most cases .
For the whole training set, Eq (1) can be re-written in its matrix form
where , , and is the incident matrix of the annotation graph .
In order to estimate the unknown parameters ( for and for ), we aim to minimise the discrepancy between the annotation and our prediction , as well as keeping the outlier estimation sparse. Note that only contains information about which pairs of instances have received votes, but not how many. The discrepancy thus needs to weighted by the number of votes received, represented by the edge weight vector . To that end, we put a weighted loss on the discrepancy and a sparsity enhancing penalty on the outlier variables. This gives us the following cost function:
and is the sparsity constraint on . With this cost function, our Unified Robust Learning to Rank (URLR) framework identifies outliers globally by integrating all local pairwise comparisons together. Note that in Eq (3), the noise term has been removed because the discrepancy is mainly caused by outliers due to their larger magnitude.
Ideally the sparsity enhancing penalty term should be a regularisation term. However, for a tractable solution, a regularisation term is used: , where is a free parameter corresponding to the weight for the regularisation term. With this penalty term, the cost function becomes convex:
where , is the diagonal matrix of and .
Setting , the problem of minimisation of the cost function in (4) can be decomposed into the following two subproblems:
Estimating the parameters of the prediction function :
Mathematically, the Moore-Penrose pseudo-inverse of is defined as , where
is the identity matrix. The scalar variableis introduced to avoid numerical instability , and typically assumes a small value222In this work, is set to .. With the the introduction of , Eq (5) becomes:
A standard solver for Eq (6) has a computational complexity, which is almost linear with respect to the size of the graph if . Faster algorithms based on the Krylov iterative and algebraic multi-grid methods  can also be used.
3.3 Outlier detection by regularisation path
From the formulations described above, it is clear that outlier detection by solving Eq (8) is the key – once the outliers are identified, the estimated can be used to substitute in Eq (5) and the estimation of the prediction function parameter becomes straightforward. Now let us focus on solving Eq (8) for outlier detection.
Note that solving Eq (8) is essentially a LASSO (Least Absolute Shrinkage and Selection Operator)  problem. For a LASSO problem, tuning the regularisation parameter is notoriously difficult [44, 45, 46, 47]. In particular, in our URLR framework, the value directly decides the ratio of outliers in the training set which is unknown. A number of methods for determining exist, but none is suitable for our formulation:
Some heuristics rules on setting the value of such as are popular in existing robust ranking models such as the M-estimator , where is a Gaussian variance set manually based on human prior knowledge. However setting a constant value independent of dataset is far from optimal because the ratio of outliers may vary for different crowdsourced datasets.
Cross validation is also not applicable here because each edge is associated with a variable and any held-out edge also has an associated unknown variable . As a result, cross validation can only optimise part of the sparse variables while leaving those for the held-out validation set undetermined.
The other alternatives e.g. Akaike information criterion (AIC) and Bayesian information criterion (BIC) are often unstable in outlier detection LASSO problems 333We found empirically that the model automatically selected by BIC or AIC failed to detect any meaningful outliers in our experiments. For details of the experiments and a discussion on the issue of determining the outlier ratio, please visit the project webpage at http://www.eecs.qmul.ac.uk/~yf300/ranking/index.html.
This inspires us to sequentially consider all available solutions for all sparse variables along the Regularisation Path (RP) by gradually decreasing the value of the regularisation parameter from to . Specifically, based on the piecewise-linearity property of LASSO, a regularisation path can be efficiently computed by the R-package “glmnet" 444http://cran.r-project.org/web/packages/glmnet/glmnet.pdf. When , the regularisation parameter will strongly penalise outlier detection: if any annotation is taken as an outlier, it will greatly increase the value of the cost function in Eq (8). When is changed from to , LASSO555For a thorough discussion from a statistical perspective, please see [49, 50, 51, 47]. will first select the variable subset accounting for the highest deviations to the observations in Eq (8). These high deviations should be assigned higher priority to represent the nonzero elements666This is related with LASSO for covariate selection in a graph. Please see  for more details. of of Eq (2), because compensates the discrepancy between annotation and prediction. Based on this idea, we can order the edge set according to which nonzero appears first when is decreased from to . In other words, if an edge whose associated outlier variable becomes nonzero at a larger
value, it has a higher probability to be an outlier. Following this order, we identify the topedge set as the annotation outliers. And its complementary set are the inliers. Therefore, the outcome of estimating using Eq (8) is a binary outlier indicator vector :
where each element indicates whether the corresponding edge is an outlier or not.
Now with the outlier indicator vector estimated using regularisation path, instead of estimating by substituting in Eq (5) with an estimated , can be computed as
where , that is, we use to ‘clean up’ before estimating .
The pseudo-code of learning our URLR model is summarised in Algorithm 1.
3.4.1 Advantage over majority voting
The proposed URLR framework identifies outliers globally by integrating all local pairwise comparisons together, in contrast to the local aggregation based majority voting. Figure 2(a) illustrates why our URLR framework is advantageous over the local majority voting method for outlier detection. Assume there are five images with five pairs of them compared three times each, and the correct global ranking order of these 5 images in terms of a specific SVP is . Figure 2(a) shows that among the five compared pairs, majority voting can successfully identify four outlier cases: , , , and , but not the fifth one . However when considered globally, it is clear that is an outlier because if we have , we can deduce . Our formulation can detect this tricky outlier. More specifically, if the estimated makes , it has a small local inconsistency cost for that minority vote edge . However, such value will be ‘propagated’ to other images by using the voting edges , , , and , which are accumulated into a much bigger global inconsistency with the annotation. This enables our model to detect as an outlier, contrary to the majority voting decision. In particular, the majority voting will introduce a loop comparison which is the well-known Condorcet’s paradox [33, 19].
We further give two more extreme cases in Figures 2(b) and (c). Due to the Condorcet’s paradox, in Figure 2(b) the estimated from majority voting, which removes , is even worse than that from all annotation pairs which at least save the correct annotation . Furthermore, Figure 2(c) shows that when each pair only receives votes in one direction, majority voting will cease to work altogether, but our URLR can still detect outliers by examining the global cost. This example thus highlights the capability of URLR in coping with extremely sparse pairwise comparison labels. In our experiments (see Section 4), the advantage of URLR over majority is validated on various SVP prediction problems.
3.4.2 Advantage over robust statistical ranking
Our framework is closely related to Huber’s theory of robust regression , which has been used for robust statistical ranking . In contrast to learning to rank, robust statistical ranking is only concerned with ranking a set of training instances by integrating their (noisy) pairwise rankings. No low-level feature representation of the instances is used as robust ranking does not aim to learn a ranking prediction function that can be applied to unseen test data. To see the connection between URLR with robust ranking, consider the Huber M-estimator  which aims to estimate the optimal global ranking for a set of training instances by minimising the following cost function:
where is the ranking score vector storing the global ranking score of each training instance
. The Huber’s loss functionis defined as
Using this loss function, when , the comparison is taken as a “good” one and penalised by an loss for Gaussian noise. Otherwise, it is regarded as a sparse outlier and penalised by an loss. It can be shown  that robust ranking with Huber’s loss is equivalent to a LASSO problem, which can been applied to joint robust ranking and outlier detection . Specifically, the global ranking of the training instances and the outliers in the pairwise rankings can be estimated as
Our URLR can be considered as a generalisation of the Huber-LASSO based robust ranking problem above. Comparing Eq (12) with Eq (3), it can be seen that the main difference between URLR and conventional robust ranking is that in URLR the cost function has the low-level feature matrix computed from the training instances, and the prediction function parameter , such that . This is because the objective of URLR is to predict SVP for unseen test data. However, URLR and robust ranking do share one thing in common – the ability to detect outliers in the training data based on a Huber-LASSO formulation. This means that, as opposed to our unified framework with feature , one could design a two-step approach for learning to rank by first identifying and removing outliers using Eq (12), followed by introducing the low-level feature matrix and prediction model parameter and estimating using Eq (9). We call this approach Huber-LASSO-FL based learning to rank which differs from URLR mainly in the way outliers are detected without considering low level features.
Next we show that there is a critical theoretical advantage of URLR over conventional Huber-LASSO in detecting outliers from the training instances. This is due to the difference in the projection space for estimating which is denoted as . To explain this point, we decompose in Eq (8
) by Singular Value Decomposition (SVD),
where with being an orthogonal basis of the column space of and an orthogonal basis of its complement. Therefore, due to the orthogonality and , we can simplify Eq (8) into
The SVD orthogonally projects onto the column space of and its complement, while is an orthogonal basis of the column space and is the orthogonal basis of its complement (i.e. the kernel space of ). With the SVD, we can now compute the outliers by solving Eq (15) which again is a LASSO problem , where outliers provide sparse approximations of projection . We can thus compare dimensions of the projection spaces of URLR and Huber-LASSO-FL:
Robust ranking based on the featureless Huber-LASSO-FL777We assume that the graph is connected, that is, ; we thus have .: to see the dimension of the projection space , i.e. the space of cyclic rankings [19, 53], we can perform a similar SVD operation and rewrite Eq (12) in the same form as Eq (15), but this time we have , and . So the dimension of for Huber-LASSO-FL is .
URLR: in contrast we have , and . So the dimension of for URLR is .
From the above analysis we can see that given a very sparse graph with , the projection space for Huber-LASSO-FL will have a dimension () too small to be effective for detecting outliers. In contrast, by exploiting a low dimensional () feature representation of the original node space, URLR can enlarge the projection space to that of dimension . Our URLR is thus able to enlarges its outlier detection projection space . As a result our URLR can better identify outliers, especially for sparse pairwise annotation graphs. In general, this advantage exists when the feature dimension is smaller than the number of training instance , and the smaller the value of , the bigger the advantage over Huber-LASSO. In practice, given a large training set we typically have . On the other hand, when the number of instances is small, and each instance is represented by a high-dimensional feature vector, we can always reduce the feature dimension using techniques such as PCA to make sure that . This theoretical advantage of URLR over conventional Huber-LASSO in outlier detection is validated experimentally in Section 4.
3.4.3 Regularisation on
It is worth mentioning that in the cost function of URLR (Eq (3)), there are two sets of variables to be estimated, and , but only one regularisation term on to enforce sparsity. When the dimensionality of (i.e. ) is high, one would expect to see a regularisation term on
(e.g. ridge regression) due tothe fact that the coefficients of highly correlated low-level features can be poorly estimated and exhibit high variance without imposing a proper size constraint on the coefficients . The reason we do not include such a regularisation term is because, as mentioned above, using URLR we need to make sure the low-level feature space dimensionality is low, which means that the dimensionality of is also low, making the regularisation term redundant. This leads to the applicability of much simpler solvers and we show empirically in the next section that satisfactory results can be obtained with this simplification.
|Dataset||No. pairs||No. img/video||Feature Dim.||No. classes|
|Image Interestingness ||2222||932 (150)||1|
|Video Interestingness ||420||1000 (60)||14|
|PubFig [54, 2]||772||557 (100)||8|
|Scene [55, 2]||2688||512 (100)||8|
|FG-Net Face Age Dataset ||–||1002||55||–|
Experiments were carried out on five benchmark datasets (see Table I) which fall into three categories: (1) experiments on estimating subjective visual properties (SVPs) that are useful on their own including image (Section 4.1) and video interestingness (Section 4.2), (2) experiments on estimating SVPs as relative attributes for visual recognition (Section 4.3), and (3) experiments on human age estimation from face images (Section 4.4). The third set of experiments can be considered as synthetic experiments – human age is not a subjective visual property although it is ambiguous and poses a problem even for humans . However, as ground truth is available, this set of experiments are designed to gain insights into how different SVP prediction models work.
4.1 Image interestingness prediction
Datasets The image interestingness dataset was first introduced in  for studying memorability. It was later re-annotated as an image interestingness dataset by . It consists of images. Each was represented as a dimensional attribute888We delete attribute features from the original feature vector in [24, 4] such as “attractive” because they are highly correlated with image interestingness. feature vector [24, 4] such as central object, unusual scene and so on. pairwise comparisons were collected by  using AMT and used as annotation. On average, each image is viewed and compared with 11.9 other images, resulting a total of 16000 pairwise labels999On average, for each labelled pair, around 80% of the annotations agree with one ranking order and 20% the other..
Settings images were randomly selected for training and the remaining for testing. All the experiments were repeated times with different random training/test splits to reduce variance. The pruning rate was set to . We also varied the number of annotated pairs used to test how well each compared method copes with increasing annotation sparsity.
Evaluation metrics For both image and video interestingness prediction, Kendall tau rank distance was employed to measure the percentage of pairwise mismatches between the predicted ranking order for each pair of test data using their prediction/ranking function scores, and the ground truth ranking provided by  and  respectively. Larger Kendall tau rank distance means lower quality of the ranking order predicted.
Competitors We compare our method (URLR) with four competitors.
Maj-Vot-1 : this method uses majority voting for outlier pruning and rankSVM for learning to rank.
Maj-Vot-2 : this method also first removes outliers by majority voting. After that, the fraction of selections by the pairwise comparisons for each data point is used as an absolute interestingness score and a regression model is then learned for prediction. Note that Maj-Vot-2 was only compared in the experiments on image and video interestingness prediction, since only these two datasets have enough dense annotations for Maj-Vot-2.
Raw: our URLR model without outlier detection, that is, all annotations are used to estimate .
Comparative results The interestingness prediction performance of the various models are evaluated while varying the amount of pairwise annotation used. The results are shown in Figure 3 (left). It shows clearly that our URLR significantly outperforms the four alternatives for a wide range of annotation density. This validates the effectiveness of our method. In particular, it can be observed that: (1) The improvement over Maj-Vot-1  and Maj-Vot-2  demonstrates the superior outlier detection ability of URLR due to global rather than local outlier detection. (2) URLR is superior to Huber-LASSO-FL because the joint outlier detection and ranking prediction framework of URLR enables the enlargement of the projection space for (see Section 3.4.2), resulting in better outlier detection performance. (3) The performance of Maj-Vot-2  is the worst among all methods compared, particularly so given sparser annotation. This is not surprising – in order to get an reliable absolute interestingness value, dozens or even hundreds of comparisons per image are required, a condition not met by this dataset. (4) The performance of Huber-LASSO-FL is also better than Maj-Vot-1 and Maj-Vot-2 suggesting even a weaker global outlier detection approach is better then the majority voting based local one. (5) Interestingly even the baseline method Raw gives a comparable result to Maj-Vot-1 and Maj-Vot-2 which suggests that just using all annotations without discrimination in a global cost function (Eq (4)) is as effective as majority voting101010One intuitive explanation for this is that given a pair of data with multiple contradictory votes, using Raw, both the correct and incorrect votes contribute to the learned model. In contrast, with Maj-Vot, one of them is eliminated, effectively amplifying the other’s contribution in comparison to Raw. When the ratio of outliers gets higher, Maj-Vot will make more mistakes in eliminating the correct votes. As a result, its performance drops to that of Raw, and eventually falls below it. .
Figure 3 (right) evaluates how the performances of URLR and Huber-LASSO-FL are affected by the pruning rate . It can be seen that the performance of URLR is improving with an increasing pruning rate. This means that our URLR can keep on detecting true positive outliers. The gap between URLR and Huber-LASSO-FL gets bigger when more comparisons are pruned showing Huber-LASSO-FL stops detecting outliers much earlier on. However, when the pruning rate is over 55%, since most outliers have been removed, inliers start to be pruned, leading to poorer performance.
Qualitative Results Some examples of outlier detection using URLR
are shown in Figure 4. It can be seen that those in the green boxes are clearly outliers and are detected correctly by our URLR. The failure cases are interesting. For example, in the bottom case, ground truth indicates that the woman sitting on a bench is more interesting than the nice beach image, whilst our URLR predicts otherwise. The odd facial appearance
The odd facial appearanceon that woman or the fact that she is holding a camera could be the reason why this image is considered to be more interesting than the otherwise more visually appealing beach image. However, it is unlikely that the features used by URLR are powerful enough to describe such fine appearance details.
4.2 Video interestingness prediction
Datasets The video interestingness dataset is the YouTube interestingness dataset introduced in . It contains categories of advertisement videos (e.g. ‘food’ and ‘digital products’), each of which has videos. annotators were asked to give complete interesting comparisons for all the videos in each category. So the original annotations are noisy but not sparse. We used bag-of-words of Scale Invariant Feature Transform (SIFT) and Mel-Frequency Cepstral Coefficient (MFCC) as the feature representation which were shown to be effective in  for predicting video interestingness.
Experimental settings Because comparing videos across different categories is not very meaningful, we followed the same settings as in  and only compared the interestingness of videos within the same category. Specifically, from each category we used videos and their paired comparisons for training and the remaining videos for testing. The experiments were repeated for rounds and the averaged results are reported.
Since MFCC and SIFT are bag-of-words features, we employed kernel to compute and combine the features. To facilitate the computation, the kernel is approximated by additive kernel of explicit feature mapping . To make the results of this dataset more comparable to those in , we used rankSVM model to replace Eq (9) as the ranking model. As in the image interestingness experiments, we used Kendal tau rank distance as the evaluation metric, while we find that the same results can be obtained if the prediction accuracy in  is used. The pruning rate was again set to .
Comparative Results Figure 5(a) compares the interestingness prediction methods given varying amounts of annotation, and Figure 5(b) shows the per category performance. The results show that all the observations we had for the image interestingness prediction experiment still hold here, and across all categories. However in general the gaps between our URLR and the alternatives are smaller as this dataset is densely annotated. In particular the performance of Huber-LASSO-FL is much closer to our URLR now. This is because the advantage of URLR over Huber-LASSO-FL is stronger when is close to . In this experiment, (1000s) is much greater than (20) and the advantage of enlarging the projection space for (see Section 3.4.2) diminishes.
Qualitative Results Some outlier detection examples are shown in Figure 6. In the two successful detection examples, the bottom videos are clearly more interesting than the top ones, because they (1) have a plot, sometimes with a twist, and (2) are accompanied by popular songs in the background and/or conversations. Note that in both cases, majority voting would consider them inliners. The failure case is a hard one: both videos have cartoon characters, some plot, some conversation, and similar music in the background. This thus corresponds to a truly ambiguous case which can go either way.
4.3 Relative attributes prediction
Datasets The PubFig  and Scene  datasets are two relative attribute datasets. PubFig contains images from people and attributes (‘smiling’, ‘round face’, etc.). Scene  consists of images from categories and attributes (‘openness’, ‘natrual’ etc.). Pairwise attribute annotation was collected by Amazon Mechanical Turk . Each pair was labelled by workers and majority vote was used in  to average the comparisons for each pair111111Thanks to the authors of  we have all the the raw pairs data before majority voting.. A total of and training images for PubFig and Scene respectively were labelled (i.e. compared with at least another image). The average number of compared pairs per attribute were and respectively, meaning most images were only compared with one or two other images. The annotations for both datasets were thus extremely sparse. GIST and colour histogram features were used for PubFig, and GIST alone for Scene. Each image also belongs to a class (different celebrities or scene types). These datasets were designed for classification, with the predicted relative attribute scores used as image representation.
Experimental Settings We evaluated two different image classification tasks: multi-class classification where samples from all classes were available for training and zero-shot transfer learning where one class was held out during training (a different class was used in each trial with the result averaged). Our experiment setting was similar to that in , except that image-level, rather than class-level pairwise comparisons were used. Two settings were used with different amounts of annotation noise:
Orig: This was the original setting with the pairwise annotations used as they were.
Orig+synth: By visual inspection, there were limited annotation outliers in these datasets, perhaps because these relative attributes are less subjective compared to interestingness. To simulate more challenging situations, we added random comparisons for each attribute, many of which would correspond to outliers. This will lead to around extra outliers.
The pruning rate was set to for the original datasets (Orig) and for the dataset with additional outliers inserted for all attributes of both datasets (Orig+synth).
Evaluation metrics For Scene and Pubfig datasets, relative attributes were very sparsely collected and their prediction performance is thus evaluated indirectly by image classification accuracy with the predicted relative attributes as image representation. Note that for image classification there is ground truth and its accuracy is clearly dependent on the relative attribute prediction accuracy. For both datasets, we employed the method in  to compute the image classification accuracy.
Comparative Results Without the ground truth of relative attribute values, different models were evaluated indirectly via image classification accuracy in Figure 7. The following observations can be made: (1) Our URLR always outperforms Huber-LASSO-FL, Maj-Vot-1 and Raw for all experiment settings. The improvement is more significant when the data contain more errors (Orig+synth). (2) The performance of other methods is in general consistent to what we observed in the image and video interestingness experiments: Huber-LASSO-FL is better than Maj-Vot-1 and Raw often gives better results than majority voting. (3) For PubFig, Maj-Vot-1  is better than Raw given more outliers, but it is not the case for Scene. This is probably because the annotators were more familiar with the celebrity faces in PubFig and hence their attributes than those in Scene. Consequently there should be more subjective/intentional errors for Scene, causing majority voting to choose wrong local ranking orders (e.g. some people are unsure how to compare the relative values of the ‘diagonal plane’ attribute for two images). These majority voting + outlier cases can only be rectified by using a global approach such as our URLR, and Huber-LASSO-FL to a lesser extent.
Qualitative Results Figure 8 gives some examples of the pruned pairs for both datasets using URLR. In the success cases, the left images were (incorrectly) annotated to have more of the attribute than the right ones. However, they are either wrong or too ambiguous to give consistent answers, and as such are detrimental to learning to rank. A number of failure cases (false positive pairs identified by URLR) are also shown. Some of them are caused by unique view point (e.g. Hugh Laurie’s mouth is not visible, so it is hard to tell who smiles more; the building and the street scene are too zoomed in compared to most other samples); others are caused by the weak feature representation, e.g. in the ‘male’ attribute example, the colour and GIST features are not discriminative enough for judging which of the two men has more ‘male’ attribute.
Running Cost Our algorithm is very efficient with a unified framework where all outliers are pruned simultaneously and the ranking function estimation has a closed form solution. Using URLR on PubFig, it took only 1 minutes to prune 240 images with 10722 comparisons and learn the ranking function for attribute prediction on a PC with four 3.3GHz CPU cores and 8GB memory.
4.4 Human age prediction from face images
In this experiment, we consider age as a subjective visual property of a face. This is partially true – for many people, given a face image predicting the person’s age can be subjective. The key difference between this and the other SVPs evaluated so far is that we do have the ground truth, i.e. the person’s age when the picture was taken. This enables us to perform in-depth evaluation of the significance of our URLR framework over the alternatives on various factors such as annotation sparsity, and outlier ratio (we now know the exact ratio). Outlier detection accuracy can also now be measured directly.
Dataset The FG-NET image age dataset121212http://www.fgnet.rsunit.com/ was employed which contains images of individuals labelled with ground truth ages ranging from to . The training set is composed of the images of randomly selected people and the rest used as the test set. All experiments were repeated times with different training/testing splits to reduce variability. Each image was represented by a dimension vector extracted by active appearance models (AAM) .
Crowdsourcing errors We used the ground truth age to generate the pairwise comparisons without any error. Errors were then synthesised according to human error patterns estimated by data collected by an online pilot study131313http://www.eecs.qmul.ac.uk/~yf300/survey4/: pairwise image comparisons from willingly participating “good” workers were collected as unintentional errors. So we assume they are not contributing random or malicious annotations. Thus the errors of these pairwise comparisons come from the natural data ambiguity. The human unintentional age error pattern was built by fitting the error rate against true age difference between collected pairs. As expected, humans are more error-prone for smaller age difference. Specifically, we fit quadratic polynomial function to model relation of age difference of two samples towards the chance of making an unintentional error. We then used this error pattern to generate unintentional errors. Intentional errors were introduced by ‘bad’ workers who provided random pairwise labels. This was easily simulated by adding random comparisons. In practice, human errors in crowdsourcing experiments can be a mixture of both types. Thus two settings were considered: Unint.: errors were generated following the estimated human unintentional error model resulting in around errors. Unint.+Int.: random comparisons were added on top of Unint., giving an error ratio of around , unless otherwise stated. Since the ground-truth age of each face image is known to us, we can give an upper bound for all the compared methods by using ground-truth age of training data to generate a set of pairwise comparisons. This outlier-free dataset is then used to learn a kernel ridge regression with Gaussian kernel. This ground-truth data trained model is denoted as GT.
Quantitative results Four experiments were conducted using different settings to show the effectiveness of our URLR method quantitatively.
(1) URLR vs. Huber-LASSO-FL. In this experiment, training images and unique comparisons were randomly sampled from the training set. Figure 9 shows that URLR and Huber-LASSO-FL improve over Raw indicating that outliers are effectively pruned using both global outlier detection methods. Both methods are robust to low error rate (Figure 9 Left: in Unint.) and are fairly close to GT, whilst the performance of URLR is significantly better than that of Huber-LASSO-FL given high error ratio (Figure 9 Right: in Unint.+Int.) because of the using low-level feature representation to increase the dimension of projection space dimension for from for Huber-LASSO-FL to for URLR (see Section 3.4.2). This result again validates our analysis that higher leads to better chance of identifying outliers correctly. It is noted that in Figure 9(Right), given outliers, the result indeed peaks when is around 25; importantly, it stays flat when up to 50% of the annotations are pruned.
(2) Comparison with Maj-Vot-1. Given the same data but each pair compared by workers (instead of ) under the Unint.+Int. error condition, Figure 10 shows that Maj-Vot-1 beats Raw. This shows that for relative dense graph, majority voting is still a good strategy of removing some outliers and improves the prediction accuracy. However, URLR outperforms Maj-Vot-1 after the pruning rate passes . This demonstrates that aggregating all paired comparisons globally for outlier pruning is more effective than aggregating them locally for each edge as done by majority voting.
(3) Effects of error ratio. We used the Unint.+Int. error model to vary the amount of random comparisons and simulate different amounts of errors in sampled graphs from training images and unique sampled pairs from the training images. The pruning rate was fixed at . Figure 11 shows that URLR remains effective even when the true error ratio reaches as high as . This demonstrates that although a sparse outlier model is assumed, our model can deal with non-sparse outliers. It also shows that URLR consistently outperforms the alternative models especially when the error/outlier ratio is high.
What are pruned and in what order? The effectiveness of the employed regularisation path method for outlier detection can be examined as decreases to produce a ranked list for all pairwise comparisons according to the outlier probability. Figure 12 shows the relationship between the pruning order (i.e. which pair is pruned first) and ground truth age difference and illustrated by examples. It can be seen that overall outliers with larger age difference tend to be pruned first. This means that even with a conservative pruning rate, obvious outliers (potentially causing more performance degradation in learning) can be reliably pruned by our model.
5 Conclusions and Future Work
We have proposed a novel unified robust learning to rank (URLR) framework for predicting subjective visual properties from images and videos. The key advantage of our method over the existing majority voting based approaches is that we can detect outliers globally by minimising a global ranking inconsistency cost. The joint outlier detection and feature based rank prediction formulation also provides our model with an advantage over the conventional robust ranking methods without features for outlier detection: it can be applied with a large number of candidates in comparison but a sparse sampling in crowdsourcing. The effectiveness of our model in comparison with state-of-the-art alternatives has been validated on the tasks of image and video interestingness prediction and predicting relative attributes for visual recognition. Its effectiveness for outlier detection has also been evaluated in depth in the human age estimation experiments.
By definition subjective visual properties (SVPs) are person-dependent. When our model is learned using pairwise labels collected from many people, we are essentially learning consensus – given a new data point the model aims to predict its SVP value that can be agreed upon by most people. However, the predicted consensual SVP value could be meaningless for a specific person when his/her taste/understanding of the SVP is completely different to that of most others. How to learn a person-specific SVP prediction model is thus part of the on-going work. Note that our model is only one of the possible solutions to inferring global ranking from pairwise comparisons. Other models exist. In particular, one widely studied alternative is the (Bradley-Terry-Luce (BTL) model [60, 61, 62]), which aggregates the ranking scores of pairwise comparisons to infer a global ranking by maximum likelihood estimation. The BTL model is introduced to describe the probabilities of the possible outcomes when individuals are judged against one another in pairs . It is primarily designed to incorporate contextual information in the global ranking model. We found that directly applying the BTL model to our SVP prediction task leads to much inferior performance because it does not explicitly detect and remove outliers. However, it is possible to integrate it into our framework to make it more robust against outliers and sparse labels whilst preserving its ability to take advantage of contextual information. Other new directions include extending the presented work to other applications where noisy pairwise labels exist, both in vision such as image denoising , iterative search and active learning of visual categories , and in other fields such as statistics and economics .
This research of Jiechao Xiong was supported in part by National Natural Science Foundation of China: 61402019, and China Postdoctoral Science Foundation: 2014M550015. The research of Yuan Yao was supported in part by National Basic Research Program of China under grant 2012CB825501 and 2015CB856000, as well as NSFC grant 61071157 and 11421110001. The research of Yanwei Fu and Tao Xiang was in part supported by a joint NSFC-Royal Society grant 1130360, IE110976 with Yuan Yao. Yuan Yao and Tao Xiang are the corresponding authours.
-  J. Donahue and K. Grauman, “Annotator rationales for visual recognition,” in ICCV, 2011.
-  A. Kovashka, D. Parikh, and K. Grauman, “Whittlesearch: Image search with relative attribute feedback,” in CVPR, 2012.
-  S. Dhar, V. Ordonez, and T. L. Berg, “High level describable attributes for predicting aesthetics and interestingness,” in CVPR, 2011.
-  M. Gygli, H. Grabner, H. Riemenschneider, F. Nater, and L. V. Gool, “The interestingness of images,” in ICCV, 2013.
-  Y.-G. Jiang, YanranWang, R. Feng, X. Xue, Y. Zheng, and H. Yang, “Understanding and predicting interestingness of videos,” in AAAI, 2013.
-  D. Parikh and K. Grauman, “Relative attributes,” in ICCV, 2011.
-  Z. Zhang, C. Wang, B. Xiao, W. Zhou, and S. Liu, “Robust relative attributes for human action recognition,” Pattern Analysis and Applications, 2013.
-  K. Chen, C. Wu, Y. Chang, and C. Lei, “Crowdsourceable QoE evalutation framework for multimedia content,” in ACM MM, 2009.
-  Y. Ma, T. Xiong, Y. Zou, and K. Wang, “Person-specific age estimation under ranking framework,” in ACM ICMR, 2011.
-  O. Chapelle and S. S. Keerthi, “Efficient algorithms for ranking with svms,” Inf. Retr., 2010.
-  X. Chen and P. N. Bennett, “Pairwise ranking aggregation in a crowdsourced setting,” in ACM International Conference on Web Search and Data Mining, 2013.
-  O. Wu, W. Hu, and J. Gao, “Learning to rank under multiple annotators,” in IJCAI, 2011.
-  C. Long, G. Hua, and A. Kapoor, “Active visual recognition with expertise estimation in crowdsourcing,” in ICCV, 2013.
-  A. Kittur, E. H. Chi, and B. Suh., “Crowdsourcing user studies with mechanical turk,” in ACM CHI, 2008.
-  A. Kovashka and K. Grauman, “Attribute adaptation for personalized image search,” in The IEEE International Conference on Computer Vision (ICCV), December 2013.
-  P. Welinder, S. Branson, S. Belongie, and P. Perona, “The multidimensional wisdom of crowds,” in NIPS, pp. 2424–2432, 2010.
V. C. Raykar, S. Yu, L. H. Zhao, A. Jerebko, C. Florin, G. H. Valadez, L. Bogoni, and L. Moy, “Supervised learning from multiple experts: Whom to trust when everyone lies a bit,” inICML, pp. 889–896, 2009.
-  J. Whitehill, T. fan Wu, J. Bergsma, J. R. Movellan, and P. L. Ruvolo, “Whose vote should count more: Optimal integration of labels from labelers of unknown expertise,” in NIPS, 2009.
-  X. Jiang, L.-H. Lim, Y. Yao, and Y. Ye, “Statistical ranking and combinatorial hodge theory,” Math. Program., 2011.
-  R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. of the Royal Statistical Society, Series B, 1996.
-  I. Gannaz, “Robust estimation and wavelet thresholding in partial linear models,” Stat. Comput., vol. 17, pp. 293–310, 2007.
-  F. L. Wauthier, N. Jojic, and M. I. Jordan, “A comparative framework for preconditioned lasso algorithms,” in Neural Information Processing Systems, 2013.
-  Y. Ke, X. Tang, , and F. Jing, “The design of high-level features for photo quality assessment,” in CVPR, 2006.
-  P. Isola, J. Xiao, A. Torralba, and A. Oliva, “What makes an image memorable?,” in CVPR, 2011.
-  F. Liu, Y. Niu, and M. Gleicher, “Using web photos for measuring video frame interestingness,” in IJCAI, 2009.
-  C. H. Lampert, H. Nickisch, and S. Harmeling, “Learning to detect unseen object classes by between-class attribute transfer,” in CVPR, 2009.
-  A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth, “Describing objects by their attributes,” in CVPR, 2009.
A. Shrivastava, S. Singh, and A. Gupta, “Constrained semi-supervised learning via attributes and comparative attributes,” inECCV, 2012.
A. Parkash and D. Parikh, “Attributes for classifier feedback,” inECCV, 2012.
-  A. Biswas and D. Parikh, “Simultaneous active learning of classifiers and attributes via relative feedback,” in CVPR, 2013.
-  A. Sorokin and D. Forsyth, “Utility data annotation with amazon mechanical turk,” in CVPR Workshops, 2008.
-  G. Patterson and J. Hays, “Sun attribute database: Discovering, annotating, and recognizing scene attributes.,” in CVPR, 2012.
-  W. V. Gehrlein, “Condorcet’s paradox,” Theory and Decision, 1983.
-  L. Liang and K. Grauman, “Beyond comparing image pairs: Setwise active learning for relative attributes,” in CVPR, 2014.
-  Q. Xu, Q. Huang, T. Jiang, B. Yan, W. Lin, and Y. Yao, “Hodgerank on random graphs for subjective video quality assessment,” IEEE TMM, 2012.
-  Q. Xu, Q. Huang, and Y. Yao, “Online crowdsourcing subjective image quality assessment,” in ACM MM, 2012.
-  M. Maire, S. X. Yu, and P. Perona, “Object detection and segmentation from joint embedding of parts and pixels,” in ICCV, 2011.
-  Z. Cao, T. Qin, T.-Y. Liu, M.-F. Tsai, and H. Li, “Learning to rank: From pairwise approach to listwise approach,” in ICML, 2007.
-  Y. Liu, B. Gao, T.-Y. Liu, Y. Zhang, Z. Ma, S. He, and H. Li, “Browserank: letting web users vote for page importance,” in ACM SIGIR, 2008.
-  Z. Sun, T. Qin, Q. Tao, and J. Wang, “Robust sparse rank learning for non-smooth ranking measures,” in ACM SIGIR, 2009.
-  Y. Fu, T. M. Hospedales, T. Xiang, S. Gong, and Y. Yao, “Interestingness prediction by robust learning to rank,” in ECCV, 2014.
-  T. Hastie, R. Tibshirani, and J. Friedman, The elements of statistical learning:Data Mining, Inference, and Prediction (2nd). Springer, 2009.
-  A. N. Hirani, K. Kalyanaraman, and S. Watts, “Least squares ranking on graphs,” arXiv:1011.1716, 2010.
-  P. J. Huber, Robust Statistics. New York: Wiley, 1981.
T. Sun and C.-H. Zhang, “Scaled sparse linear regression,”Biometrika, vol. 99, no. 4, pp. 879–898, 2012.
-  A. Belloni, V. Chernozhukov, and L. Wang, “Pivotal recovery of sparse signals via conic programming,” Biometrika, vol. 98, pp. 791–806, 2011.
-  Y. She and A. B. Owen, “Outlier detection using nonconvex penalized regression,” Journal of American Statistical Association, 2011.
-  J. Friedman, T. Hastie, and R. Tibshirani, “Regularization paths for generalized linear models via coordinate descent,” Journal of Statistical Software, vol. 33, no. 1, pp. 1–22, 2010.
-  J. Fan and R. Li, “Variable selection via nonconcave penalized likelihood and its oracle properties,” JASA, 2001.
-  J. Fan, R. Tang, and X. Shi, “Partial consistency with sparse incidental parameters,” arXiv:1210.6950, 2012.
-  B. Efron, T. Hastie, I. Johnstone, and R. Tibshirani, “Least angle regression,” Annals of Statistics, 2004.
-  N. Meinshausen and P. Bühlmann, “High-dimensional graphs and variable selection with the lasso,” Ann. Statist., 2006.
-  Q. Xu, J. Xiong, Q. Huang, and Y. Yao, “Robust evaluation for quality of experience in crowdsourcing,” in ACM MM, 2013.
-  N. Kumar, A. C. Berg, P. N. Belhumeur, and S. K. Nayar, “Attribute and simile classifiers for face verification,” in ICCV, 2009.
-  A. Oliva and A. Torralba., “Modeling the shape of the scene: Aholistic representation of the spatial envelope,” IJCV, vol. 42, 2001.
-  Y. Fu, G. Guo, and T. Huang, “Age synthesis and estimation via faces: A survey,” TPAMI, 2010.
-  S. Mika, B. Scholkopf, A. Smola, K.-R. Muller, M. Scholz, and G. Ratsch, “Kernel PCA and de-noising in feature spaces,” in NIPS, pp. 536–542, 1999.
-  A. Vedaldi and A. Zisserman, “Efficient additive kernels via explicit feature maps,” in IEEE TPAMI, 2011.
-  B. Carterette, “On rank correlation and the distance between rankings,” in ACM SIGIR, 2009.
-  R. Hunter, “Mm algorithms for generalized bradley-terry models,” The Annals of Statistics, vol. 32, p. 2004, 2004.
H. Azari Soufiani, W. Chen, D. C. Parkes, and L. Xia, “Generalized Method-of-Moments for Rank Aggregation,” inNIPS, 2013.
F. Caron and A. Doucet, “Efficient bayesian inference for generalized bradley-terry models,” 2012.
-  S. X. Yu, “Angular embedding: A robust quadratic criterion,” TPAMI, 2012.
-  T.-K. Huang, R. C. Weng, and C.-J. Lin, “Generalized bradley-terry models and multi-class probability estimates,” The Journal of Machine Learning Research, vol. 7, pp. 85–115, 2006.
-  J. A. Tropp and A. C. Gilbert, “Signal recovery from random measurements via orthogonal matching pursuit,” IEEE Information theory, 2007.
-  M.G.Kendall and J.D.Gibbons, Rank correlation methods. Ox, 1990.
-  P. Isola, D. Parikh, A. Torralba, and A. Oliva, “Understanding the intrinsic memorability of images,” in NIPS, 2011.
Thanks for the excellent questions from the anonymous reviewers of our TPAMI submission. In answering their questions, we found some details and insights of our framework which have been overlooked before. Due to the page limits of our journal version, we use this document to further explain the details and insights and help our readers better understand our work.
Further, the proposed approach doesn’t seem to truly get to the bottom of why subjective properties are tricky namely that two people might actually have a different understanding of the property. While the authors do refer to such possible disagreements in the introduction, the proposed method doesn’t seem to consider this possibility. In other words, how does it make sense to consider a single global order when such an order might be unattainable since person A’s "interestingness" will differ from person B’s?
This is a very good question. Indeed, since the properties are subjective, they are by definition person-dependent. However, in most applications when we learn a SVP prediction model using pairwise labels collected from many different annotators, we are modeling consensus. In other words, the model essentially aggregates the understandings of different people regarding a certain SVP so that the predicted SVP for an unseen data point can be agreed upon by most people. For example, in the case of video interestingness, YouTube may want to predict the interestingness of a newly uploaded video so as to decide whether or not to promote it. Such a prediction obviously needs to be based on consensus from the majority of the YouTube viewers regarding what defines interestingness. However, collecting consensus can be expensive; the proposed model in this paper thus aims to infer the consensus from as few labels as possible.
It is also true that for a specific person, he/she would prefer a SVP prediction model that is tailor-made for his/her own understanding of the SVP, i.e. a person-specific prediction model. Such a model needs to be learned using his/her pairwise labels only. For example, YouTube could recommend different videos for different registered users when they log in, if they provide some pairwise video interestingness labels for learning such a model (at present, this is done based on some simple rules from the viewing history of the user). This also has its own problem - it is much harder to collect enough labels from a single person only to learn the prediction model. There are solutions, e.g. categorising the users into different groups so that the labels from people of the same group can be shared. However this is beyond the scope of this paper and is being considered as part of ongoing work.
We have provide a discussion on this problem in Section 5 in the revised manuscript (Page 14).
It feels a little bit unsatisfying that the method requires we pick a fixed ratio of outliers. This would be more ok if the ratio can be automatically computed from the data somehow.
Indeed, the pruning rate is a free parameter of the proposed model (in fact, the only free parameter) that has to set manually. As discussed in the beginning of Section 3.3, most existing outlier detection algorithms have a similar free parameter determining how aggressive the algorithm needs to be for pruning outliers. Automated model selection critieria such as BIC and AIC could be considered. However, as pointed out by , they are often unstable for the outlier detection problem with pairwise labels. We have carried out experiments to show that when BIC or AIC is employed, the selected model failed to detect meaningful outliers. Since a related comment is given by Reviewer 3, please refer to the Response Point 2 to Reviewer 3 for detailed experiment results and analysis on the alternative outlier detection methods including BIC. It is also worth pointing out that our results on the effect of the pruning rate show that the proposed model remains effective given a wide range of pruning rate values (see Fig. 3, 5, 9 and 10).
We have now added a footnote in Section 3.3 to discuss why an automated model selection criterion such as BIC is not adopted.
I think cases of Raw performing similarly or better than MajVot1/ 2 should be explained in a little more detail, i.e. an intuition for such outcomes should be given.
Thanks for the suggestion. Indeed, our results on both image and video interestingness experiments show that Raw performs similarly to majority voting. There is an intuitive explanation for that. When a pair of data points A and B receive multiple votes/labels of different pairwise orders/ranks, these multiple labels are converted into a single label corresponding to the order that receives the most votes. Since only one of the two orders is correct (either A>B or B>A), there are two possibilities: the majority voted label is correct, or incorrect, i.e. an outlier. In comparison, using Raw, all votes count, so the outlying votes would certainly having a negative effect on the learned prediction model, so would the correct votes/labels. Now let us consider which method is better. The answer is it depends on the outlier/error ratio of the labels. If the ratio is very low, majority voting will get rid of almost all the outlying votes; MajVot would thus be advantageous over Raw which can still feel the negative effects of the outliers. However, when the ratio gets bigger, it becomes possible that the outlying label becomes the winning vote. For example, if A>B is correct, and received 2 votes and A<B is incorrect and received 3 outlying votes. Using Raw, those 2 correct votes still contribute positively to the model, whilst using MajVot, their contribution disappears and the negative impact of the outlying votes is amplified. Therefore, one expects that when MajVot makes more and more mistakes, its performance will get closer to that of Raw, until it reaches a tipping point where Raw starts to get ahead.
We have added a brief discussion on this in Section 4.1. on Page 9.
In Figure 3, why does the Kendall tau distance start to increase as the pruning rate increases, after 55% for URLR?
Higher Kendall tau distance means worse prediction. Figure 3 (right) thus shows that our URLR’s performance is improved when more and more outliers are pruned in the beginning; then after more than 55% of pairs are pruned, its performance starts to decrease. This result is expected: at low pruning rates, most of the pruned pairs are outliers; the model therefore benefits. Since the percentage of outliers would almost certainly be lower than 50%, when the pruning rate reaches 55%, most of the outliers have been removed, and the algorithm start to remove the correctly labelled pairs. With less and less correct labels available to learn the model, the performance naturally would decrease - when pruning rate gets close to 100%, it would not be possible to learn a meaningful model; the Kendall tau distance would thus shoot up.
We have now added a sentence on Page 10 to give an explanation to this phenomenon.
Page 2 line 44 “For example, Figure 1 … ” the authors try to argue that examples shown in Figure 1 are outliers. I don’t quite agree. Authors are trying to study subjective attributes. These are good examples of subjective versus objective attributes. This doesn’t seem to be about outliers vs. not. In fact, one source of outliers other than malicious workers is global (in)consistency, which is not mentioned here. The authors could draw from the concrete example of Figure 2.
This is a very good point. It is certainly worthwhile to clarify the definition of outlier in the context of subjective visual property (SVP). In particular, since by definition a SVP is subjective, defining outlier, even making the attempt to predict SVP is self-contradictory - one man’s meat is another man’s poison. However, there is certainly a need for learning a SVP prediction model, hence this paper. This is because when we learn the model from labels collected from many people, we essentially aim to learn the consensus, i.e. what most people would agree on (please see our Response Points 1 for more discussion on this). Therefore, Figure 1(a) can still be used to illustrate this outlier issue in SVP annotation, that is, you may have most of the annotators growing up watching Sesame Street thus consciously or subconsciously consider the Cookie Monster to be more interesting than the Monkey King; their pairwise labels/votes thus represent the consensus. In contrast, one annotator who is familiar with the stories in Journey to the West may choose the opposite; his/her label is thus an outlier under the consensus. We have reworded the relevant text on Page 2 to avoid confusion.
A baseline to compare to might be to feed all individual constraints (without majority vote) to a rankSVM. SVMs already allow for some slack. So I would be curious to know if that takes care of some of the outliers already.
Thanks. In fact, we do have one set of results on this. Specifically, in Sec 4.2 “Video interestingness prediction”, as explained under “Experimental settings”, we employed rankSVM model to replace Eq (9). Therefore, the model denoted as ’Raw’ in this experiment is exactly the suggested baseline of feeding all constraints to a rankSVM. As shown in Fig. 5(a), the model is at par with Maj-Vot-1 but worse then the two global outlier detection methods Huber-LASSO-FL and our URLR. This result suggests that rankSVM does have some ability to cope with outliers. However, we are not sure this is due to the slack variables of rankSVM. This is because the slack variables are introduced to account for data noise [ranking_books] which is different from the outliers in the pairwise data.
For the scene and pubfig image dataset, the relative attribute prediction performance can only be evaluated indirectly by image classification accuracy with the predicted relative attributes as image representation." > Why is that? Can’t you compute attribute prediction performance on a held out set of annotated pairs? Or is the concern that since the pairs may be noisily annotated, one can not think of them as GT? But is that not an issue with interestingness then? Please clarify in rebuttal.
Thanks for this question. We stated in footnote 9 that “Collecting ground truth for subjective visual properties is always problematic. Recent statistical theories ,  suggest that the dense human annotations can give a reasonable approximation of ground truth for pairwise ranking. This is how the ground truth pairwise rankings provided in  and  were collected.” So for image and video interestingness as well as the age dataset, (dense) enough pairwise comparisons are available to give a reasonable approximation of the groundtruth. However, this is not the case for scene and pubfig image dataset: the collected pairs are much more sparse and cannot be used as an approximation to the groundtruth. In short, it is because they are too sparse rather than too noisy.
In contrast, the indirect evaluation metric of downstream classification accuracy has clear unambiguous groundtruth, and directly depends on relative attribute prediction accuracy. So this evaluation is preferred.
Related Work: The Bradley-Terry-Luce (BTL) model is the standard model for computing a global ranking from pairwise labels. It should be mentioned in the related work. See  or Hunter, D. R. (2004). MM algorithms for generalized BradleyTerry models. Annals of Statistics. Experiments: I would expect additional comparisons to state-of-the-art (BTL or SVM-rank aggregation ). In particular the Bradley-Terry-Luce (BTL) model is extremely widely used and more robust to noise than LASSO based approaches . E.g. "Generalized Method-of-Moments for Rank Aggregation" or "Efficient Bayesian Inference for Generalized Bradley-Terry Models" provide code for inference in BTL models. Such a method leads to a global ranking, which could be used to train an SVM. Alternatively, it can be used to find pairwise rankings that disagree with the obtained global ranking. These could be removed as outliers and a rank-SVM trained from the remaining pairwise labels. Such an experiment should be included as an additional state-of-the-art comparison in the updated version of the manuscript.
Thanks for the suggestion. Indeed, the Bradley-Terry-Luce (BTL) model is a very relevant global ranking model. We have now studied it carefully and made connections to the proposal URLR model. We also carried out new experiments to evaluate the BLT model for our Subjective Visual Property (SVP) prediction task.
More specifically, the BTL model is a probabilistic model that aggregates the ranking scores of pairwise comparisons to infer a global ranking by maximum likelihood estimation. It is closely related to the proposed global ranking model; yet it also has some vital differences. Let’s first look at the connection. The main pairwise ranking model of Huber-LASSO used in this paper is a linear model (see Eq (10) and Eq (12)), which is
In statistics and psychology [19, 64, 51, rank_analysis_bradleyTerry], such a linear model can be extended to a family of generalised linear models when only binary comparisons are available for each pair , i.e. either is preferred to or vice versa. In these generalised linear models, one assumes that the probability of pairwise preference is fully determined by a linear ranking/rating function in the following,
can be chosen as any symmetric cumulated distributed function.
Different choices of lead to different generalised linear models. In particular, two choices are worth mentioning here:
This model is equivalent to use if is preferred to and otherwise in linear model. This model is used in this work to derive our URLR model.
Bradley-Terry-Luce (BTL) model,
So by now, it is clear that both our URLR and BTL generalise the linear model in Huber-LASSO. They differ in the choice of the symmetric cumulated distributed function .
Although both of them are generalised from the same linear model, they are developed for very different purposes. The BTL model is introduced to describe the probabilities of the possible outcomes when individuals are judged against one another in pairs [60, rank_analysis_bradleyTerry]. It is primarily designed to incorporate contextual information in the global ranking model. For instance, in sports applications, it can be used to account for the home-field advantage and ties situations [64, 62]. In contrast, our framework tries to detected outliers in the pairwise comparisons and cope with the sparse labels. Consequently, from Eq (1) onwards, we introduce the outlier variable to model the outiers explicitly and introduce low-level feature variable to enhance our model’s ability to detect outliers given sparse labels. None of these is in the BLT model, which means that it may not be suitable given sparse pairwise comparisons with outliers.
To verify this, we took the suggestion by Reviewer 3 and employed the matlab codes from the website of  "Efficient Bayesian Inference for Generalized BradleyÂTerry Models" to carry out experiments. The results on image interestingness prediction are compared in Fig 13. It shows that the performance of BTL is much worse than the other alternatives. Similar results were obtained on video interestingness prediction and age estimation.
As explained above, it is actually not fair to compare the BTL model to the other models because BTL was not designed for outlier detection and could not cope with the amount of outliers and the level of spareness in our SVP data. We therefore decide not to include the new results in the revised manuscript. However, from our analysis above, it is also clear that we could use the BTL model (Eq (18)) to generalise the linear model in place of the uniform model, and use it in our outlier detection framework. In this way, we can have the better of both worlds: the ability of BTL to incorporate contextual information such as the home-filed advantage in sports can also be taken advantage of in our framework whilst preserving our model’s strength on robustness against outliers and sparse labels. However, this is probably beyond the scope of this paper and is better left to the future work. In the revised manuscript, we have now added the following paragraph in Section 5, where we discuss that BTL is an alternative model that can be integrated into our framework as part of the future work.
“Note that our model is only one of the possible solutions to inferring global ranking from pairwise comparisons. In particular, one widely studied alternative is the (Bradley-Terry-Luce (BTL) model [61,62,63], which aggregates the ranking scores of pairwise comparisons to infer a global ranking by maximum likelihood estimation. The BTL model is introduced to describe the probabilities of the possible outcomes when individuals are judged against one another in pairs . It is primarily designed to incorporate contextual information in the global ranking model. We found that directly applying the BTL model to our SVP prediction task leads to much inferior performance because it does not explicitly detect and remove outliers. However, it is possible to integrate it into our framework to make it more robust against outliers and sparse labels whilst preserving its ability to take advantage of contextual information.”
3.3 Regularization path. On the one hand the authors say that "Setting a constant λ value independent of dataset is far from optimal because the ratio of outliers may vary for different crowdsourced datasets", but using the regularization path this is exactly what is done in the end. It is true that the experiments show that the proposed method is fairly robust w.r.t. the outlier ratio. Nonetheless, I would like to see an experiment using a (modified) BIC for selecting the outlier ratio. This would be a valuable extension over the ECCV work.
Thanks. As discussed in the beginning of Section 3.3, most existing outlier detection algorithms have a similar free parameter as λ to determine how aggressive the algorithm needs to be for pruning outliers. Automated model selection critieria such as BIC and AIC could be considered. However, as pointed out by , they are often unstable for the outlier detection problem with pairwise labels.
We have evaluated alternative methods including the modified BIC and AIC for image and video interestingness prediction. The results suggest those automated models such as AIC and BIC failed to identify any outliers - they prefer the model that include all input pairwise comparisons. To find out why it is the case, we carried out a controlled experiment using synthetic data to investigate how different factors affect the performance of different methods for determining the outlier ratio. Specifically, we compare , BIC and with our Regularization Path model.
Experiments design. we use a complete graph with nodes. Our framework is simplified into the following ranking model,
Let , and . We simulate the outlier pairs by randomly sampling, that is, each pair’s true ranking is reversed (i.e. becoming an outlier/error) with a probability which will determine the outlier ratio. The magnitude of outliers in relation to that of the noise is another factor which could potentially affect the performance of different methods on outlier detection. So we define the outlier-noise-ratio , where in our experiment and is varied in our experiment to give different ONR values.
Evaluation protocols and results. We first compare three methods that require the manual setting of a free parameter corresponding to the outlier ratio. These include our formulation (Eq (8)) with Regularization Path (i.e. the proposed model), IPOD hard-threshold 141414Strictly speaking, IPOD hard-threshold is not a Lasso solver, since it replaced the soft-thresholding with hard-thresholding. However, for comparison convenience, we still compare it with our RP. with Regularization Path, and our formulation with orthogonal matching pursuit . Using our model with Regularization Path, is decreased from to 0 and the graph edges are order according to how likely it corresponds to an outlier. The top edge set are detected as outliers. By varying
, ROC (receiver-operating-characteristic) curve can be plotted and AUC (area under the curve) is computed. Similarly, IPOD hard-threshold can also be solved using the same Regularization Path strategy. And orthogonal matching pursuit can be used to solve our formulation for outlier detection in place of Regularization Path. As shown in Figure14, the results of our formulation with Regularization Path are consistently better than those of IPOD hard-threshold + Regularization Path and our formulation + orthogonal matching pursuit. Specifically, it shows that (1) when there are small portions of outliers, all the methods can reliably prune most of outliers; (2) in all experiments, IPOD-hard threshold and orthogonal matching pursuit have similar performance, whilst our formulation + Regularization Path is consistently better than the other alternatives, especially when there are large portions of outliers (high values of ); (3) the higher the ONR, the better performance of outlier detection for all three methods.
ONR=4 ONR=5 ONR=6 ONR=7 ONR=8 ONR=9 p=0.1 0.002/0 0.494/0.012 1/0.003 1/0.026 1/0.025 1/0.031 p=0.2 0/0 0/0 0.3/0.016 0.9/0.05 1/0.064 1/0.037 p=0.3 0/0 0/0 0/0 0/0 0/0 0.5/0.06 p=0.4 0/0 0/0 0/0 0/0 0/0 0/0 p=0.5 0/0 0/0 0/0 0/0 0/0 0/0 p=0.6 0/0 0/0 0/0 0/0 0/0 0/0 Table II: The outlier detection results of our formulation + BIC. The results are presented as TPR/FPR. The error probability and ONR are: and respectively.
In contrast, BIC utilises the relative quality and likelihood functions of statistical models themselves to determine a fixed . Therefore, the true positive rate (TPR) and false positive rate (FPR) for BIC are reported. The results are listed in Table II. It shows that when using our formulation with BIC, only when there are very small portions of outliers and the outlier-noise-ratio is extremely high, BIC can reliably prune most of outliers. Otherwise, it tends to consider all pairs inliers. As mentioned above, using BIC in place of Regularization Path also leads to no outliers being pruned in our SVP prediction experiments. This thus suggests that the real outlier ratio (roughly corresponds to =0.2, see Response Point 10 to Reviewer 2) and/or outlier-noise-ratio (ONR) are too high for BIC to work.
Due to the space constraint, we could not include all these results and analysis in the revised manuscript. On Page 6, we have now added a footnote (Footnote 3) to refer the readers to find additional results and discussion on this outlier ratio problem in the project webpage at http://www.eecs.qmul.ac.uk/~yf300/ranking/index.html.
Page 9, Col. 2, Line 52: The authors talk about global image features (GIST), but Page 8, Line 45 indicates that the ground truth annotations such as “central object”, etc. were used. Using the complete ground truth annotation seems to be problematic, as it also contains an attribute "is interesting" and others such as "is aesthetic" and "is unusual". When using this ground truth, I believe such labels should be excluded and only content attributes used. (such as: indooroutdoor, contains a person, etc.).
Thanks for the suggestion. We have updated this experiment as suggested. Specifically, we first examined how each of the 932 attribute features are correlated to the groundtruth interestness value of each image. Figure 15. shows that (1) only small number of these attribute features have strong correlation with the interestingness value. (2) the histogram of kendall tau correlations151515Note that here, we employ kendall tau correlation rather than the Spearman correlation (Spearman correlation of “is interesting” vs. groundtruth is 0.63 as reported in ) since Spearman correlation is much more sensitive to error and discrepancies in data and Kendall tau correlation  generally have better statistical properties. of all features is roughly Gaussian as shown in Fig. 15(right).
attribute pleasant_scene attractive memorable is_aesthetic is_interesting on_post-card buy_painting hang_on_wall corr -0.4060 -0.4273 -0.4618 0.4487 0.4715 0.4767 0.4085 0.4209 Table III: The pruned attribute features.
So as suggested, for more fair comparisons, we remove the attribute features  whose kendall tau correlations are higher than 0.4 or lower than -0.4. This will lead to deletion the features listed in Table III. These pruned features include those suggested by Reviewer 3 (“is_interesting” and “is aesthetic”). but not the “unusual” attribute feature which has a low correlation value of -0.0226.
We repeat the image interestingness experiments with the updated features. It is noticed that this has little effect on the results (still within the variances).