Training large deep neural networks requires massive amounts of labeled training data. This fact hampers their application to domains where training data is scarce and the process of collecting new datasets is laborious and/or expensive. Recently, self-supervised learning has received attention because it offers an alternative to collecting labeled datasets. Self-supervised learning is based on the idea of using an auxiliary task (different, but related to the original supervised task) for which data is freely available and no annotation is required. As a consequence, self-supervised learning can be much more scalable. In
the self-supervised task is estimating the relative location of patches in images. Training on this task allows the network to learn features discriminative for semantic concepts. Other self-supervised tasks include generating color images from gray scale images and vice versa[2, 3], recovering a whole patch from the surrounding pixels by inpainting , and learning from equivalence relations .
In this paper, we investigate the use of ranking as a self-supervised auxiliary task. In particular we consider regression problems in computer vision for which it is easy to obtain ranked data automatically from unlabeled data. By ranked data we mean that we know that for some samples the regression output is known to be larger (or smaller) than for some others. In these cases the unlabeled data, converted into ranked subsets of data, can be added in a multi-task sense during training by minimizing an additional ranking loss. The main advantage of our approach is that it allows adding large amounts of unlabeled data to the training dataset, and as a results train better deep neural networks. In addition, we show that the ranked subsets of images can be exploited to performing active learning by identifying which images, when labeled, will result in the largest improvement of performance of the learning algorithm (here a neural network). We consider two specific computer vision regression problems to demonstrate the advantages of learning from ranked data: Image Quality Assessment (IQA) and crowd counting. In Fig. 1 we give an overview of the general approach to use labeled and unlabeled (but ranked) data in a multi-task network.
The first regression problem we consider is No-Reference Image Quality Assesment (NR-IQA), where the task is to predict the perceptual quality of images without using the undistorted image (also called the reference image). This research field has also seen large improvements in recent years due to the advent of Convolutional Neural Networks (CNNs)[6, 7, 8]. The main problems these papers had to address is the lack of large datasets for IQA. However, the annotation process for IQA image datasets requires multiple human annotations for every image, and thus the collection process is extremely labor-intensive and costly. As a result, most available IQA datasets are too small to be effective for training CNNs. We show how to automatically generate rankings for the task of IQA, and how these rankings can be used to improve the training of CNNs for NR-IQA.
The second regression problem we consider is crowd counting. Crowd counting is a daunting problem because of perspective distortion, clutter, occlusion, non-uniform distribution of people, complex illumination, scale variation, and a host of other scene-incidental imaging conditions. Techniques for crowd counting have also seen improvement recently due to the use of CNNs. These recent approaches include scale-aware regression models, multi-column CNNs , and switching networks . As with most CNN architectures, however, these person counting and crowd density estimation techniques are highly data-driven. Even modestly deep architectures for visual recognition require massive amounts of labeled training data for learning. For person counting, the labeling burden is even more onerous than usual. Training data for person counting requires that each individual person be meticulously labeled in training images. It is for this reason that person counting and crowd density estimation datasets tend to have only a few hundred images available for training. As a consequence, the ability to train these sophisticated CNN-based models suffers. We show how ranked sets of images can be generated for the task of crowd counting, and how these can be exploited to train deep networks.
In this work we study the use of ranking as a self-supervised proxy task to leverage unlabeled data and improve the training of deep networks for regression problems. The contributions of the paper are:
We show that ranking tasks can be used as self-supervised proxy tasks, and how this can be exploited to leverage unlabeled data for applications suffering from a shortage of labeled data.
We show that the ranking task on unlabeled data can be exploited as an active learning strategy to determine which images should be labeled to improve the performance of the network.
We demonstrate the advantages of the above contributes on two applications: Image Quality Assessment (IQA) and crowd counting. On both tasks we show how to effectively leverage unlabeled data and that this significantly improves performance over the state-of-the-art.
This paper is an extension of our previous work on No-reference IQA  and crowd counting . Beyond these previous works, here we propose a general framework for training from self-supervised rankings. In addition, we apply the multi-task approach from  to the IQA problem, improving the results of our previous IQA method. We also include results on two new datasets for crowd counting, UCSD and WorldExpo’10. Finally, we show that self-supervised ranking tasks can also be used as a criterion for active learning. In this case the aim is to select which images to label from a large pool of unlabeled data. In experiments we show that this can significantly reduce the required labeling effort.
This paper is organized as follows. In the next section we review work from the literature related to our work. In section 3 we describe our general learning to rank framework for self-supervised learning. In section 4 we describe how to automatically generate rankings for image quality assessment. In section 5 we show how the proposed framework can be applied to the problem of crowd counting. In sections 6.1 and 6.2 we give extensive experimental evaluations for IQA and crowd counting, respectively. We conclude in section 7 with a discussion of our contributions and some indications of potential future research lines.
2 Related work
In this section we review work from the literature on learning from rankings, active learning, image quality assessment, and crowd counting.
2.1 Learning from rankings
Several works have studied how to learn to rank, and they focus on learning a ranking function from ground-truth rankings [17, 18]. They learn a ranking function from ground-truth rankings by minimizing a ranking loss . This function can then be applied to rank test objects. The authors of 
adapt the Stochastic Gradient Descent method to perform pairwise learning to rank. This has been successfully applied to large datasets. However, these approaches are very different from ours in which we aim to learnfrom rankings.
In a recent paper 
a method is proposed where the self-supervised task is to learn to count. The authors propose two proxy tasks – scaling and tiling – which guide self-supervised training. The network learn to count visual primitives in image regions. It is self-supervised by the fact that the number of visual primitives is not expected to change under scaling, and that the sum of all visual primitives in individual tiles should equal the total number of visual primitives in the whole image. Unlike our approach, they do not consider rankings of regions and their counts are typically very low (several image primitives). Also, their final tasks do not involve counting but rather unsupervised learning of features for object recognition.
2.2 Active learning
Active learning  is a machine learning procedure that reduces the cost of annotation by actively selecting the best samples to label among the abundantly available unlabeled data. Active learning is well-motivated in many modern machine learning problems where data may be abundant, but labels are scarce or expensive to obtain. Since massive amounts of data are required to train deep neural networks, informative samples are more valuable to annotate instead of annotating randomly picked samples. Active learning has been explored in many applications such as image classification , object detection  and image segmentation .
There are several ways to approach the active learning problem. Uncertainty sampling  is the simplest and most commonly used that samples the instance whose prediction is least confident. Margin sampling 
aims to correct for the shortcomings of uncertainty sampling by examining the difference between second and first most likely labels for candidate samples. A more general strategy considers all prediction probabilities using entropy. Expected Model Change is a metric that measures change in gradient by calculating the length as an expectation over the possible labelings. However, these approaches only consider the uncertainty of instances and ignore the representatives of the underlying distribution. The method proposed in  addresses this issue by computing the similarity between the candidate instance and all other samples in the training set. Note that these active learning approaches are more about classification problems, while we are primarily interested in regression.
In this work we show how to apply active learning to image quality assessment and crowd counting – two tasks for which image annotation is expensive. Instead of using the above approaches, we measure the informativeness of unlabeled instances via mistakes made by the network on a self-supervised ranking proxy task. Performance on this proxy task is guaranteed to be consistent with the main task, and our hypothesis is that the instances with more mistakes on the proxy task vary more from the training set. Training on such instances should allows us to increase the generalizing capacity of neural networks.
2.3 Image Quality Assessment
We briefly review the IQA literature related to our approach. We focus on recent deep learning based methods for distortion-generic, No-reference IQA since it is more generally applicable than the other IQA research lines.
In recent years several works have used deep learning for NR-IQA [27, 7, 8]. One of the main drawbacks of deep networks is the need for large labeled datasets, which are currently not available for NR-IQA research. To address this problem Kang et al.  consider small patches rather than images, thereby greatly augmenting the number of training examples. The authors of [28, 8] follow the same pipeline. In  the authors design a multi-task CNN to learn the type of distortions and image quality simultaneously. Bianco at al.  propose to use a pre-trained network to mitigate the lack of training data. They extract features from a pre-trained model fine-tuned on an IQA dataset. These features are then used to train an SVR model to map features to IQA scores.
There are other works which, like us, apply rankings in the context of NR-IQA. Gao et al.  generate pairs in the dataset itself by using the ground truth scores with a threshold and combine different hand-crafted features to represent image pairs from the IQA dataset. Xu et al.  propose training a specific model for each distortion type in a multi-task learning model. Instead of using the ground truth in the dataset, they learn a ranking function. The most relevant work is from Ma et al. , which was published concurrently with our work on NR-IQA . Like us, they use learning-to-rank to deal with the extremely limited ground truth data for training. However, they generate pairing data by using other Full-reference IQA methods with a threshold like in , while our approach does not require other methods and operates in a self-supervised manner. Another difference is that we use a knowledge transfer technique for further fine-tuning on the target dataset with the same network, which can be also learned in a multi-task, end-to-end way, while they train an independent regression model based on the feature representations to perform the prediction. In addition to these differences, we show results on a larger number of distortion types instead of only working on the four main distortions .
In our paper, we propose a radically different approach to address the lack of training data: we use a large number of automatically generated rankings of image quality to train a deep network. This allows us to train much deeper and wider networks than other methods in NR-IQA which train directly on absolute IQA data.
2.4 Crowd counting
We focus on deep learning methods for crowd counting in still images. For a more complete review of the literature on crowd counting, including classical approaches, we refer the reader to .
As introduced in the review of 
, CNN-based approaches can be classified into different categories based on the properties of the CNN. Basic CNNs incorporate only basic CNN layers in their networks. The approaches in[33, 34] use the AlexNet network  to map from crowd scene patches to global number of people by changing the output of AlexNet from 1000 to 1. The resulting network can be trained end-to-end. Due to the large variations of density in different images, recent methods have focused on scale-awareness. The method proposed in  trains a multi-column based architecture (MCNN) to capture the different densities by using different sizes of kernels in the network. Similarly, the authors of  propose the Hydra-CNN architecture that takes different resolutions of patches as inputs and has multiple output layers (heads) which are combined in the end. Most recently, in  the authors propose a switching CNN that can select an optimal head instead of combining the information from all network heads. Finally, context-aware models are networks that can learn from the context of images. In [33, 36] the authors propose to classify images or patches into one of five classes: very high density, high density, medium density, low density and very low density. However, the definition of these five classes varies across datasets and must be carefully chosen using knowledge of the statistics of each dataset.
Although CNN-based methods have achieved great success in crowd counting, due to lack of labeled data it is still challenging to train deep CNNs without over-fitting. The authors of  propose to learn density map and global counting in an alternating sequence to obtain better local optima. The method in  uses side information like ground-truth camera angle and height to help the network to learn. However, this side information is expensive to obtain and is not available in most existing crowd counting datasets.
There is some interesting recent work on CNNs for crowd counting. CSRNet  consists of two components: a convolutional neural network as 2D feature extractor and a dilated CNN for estimating a density map to yield larger receptive fields and to replace pooling operations. DecideNet  starts by estimating the crowd density using detection and regression separately. It then assesses the reliability of these two estimates with an attention module. Shen et al.  propose using a U-net structure to generate a high quality density map with an adversarial loss. Shi et al.  formulate a single ConvNet as ensemble learning. Marsden et al.  adapt object counting models to new visual domains like cell counting and penguins counting. Both works [44, 45] propose generating a high resolution density map. Ierees et al.  propose solving the problems of counting, density map estimation and localization simultaneously. Laradji et al.  propose a detection-based method that does not need to estimate the size and shape of the objects.
In our paper, we show how a large number of unlabeled crowd data can improve the training of crowd counting networks. We automatically generate rankings from the unlabeled images, which are used during the training process in the self-supervised proxy task.
3 Learning from rankings
In this section we lay out a general framework for our approach, then describe how we use a Siamese network architecture to learn from rankings. In section 3.3 we show how backpropagation for training Siamese networks from ranked samples can be made significantly more efficient. Finally, in section 3.4 we show how the ranking proxy task can be used as an active learning algorithm to identify which are the most important images to label first.
3.1 Ranking as a self-supervised proxy task
Regression problems consist of finding a mapping function between input variables and a continuous output variable. It is a vital area of research in machine learning, and many important problems in computer vision are regression problems. The mapping function is typically learned from a training dataset of labelled data which consists of pairs of input and output variables. The complexity of the mapping function that can be learned is limited by the number of labeled examples in the training set. In this article, we are interested in using deep convolutional neural networks (CNNs) as mapping functions, and images as input data.
For some regression problems it can be easy to obtain a ranked dataset. Such a dataset contains relative information between pairs of input examples, describing which of the two is larger. For image quality assessment (IQA) it is easy to generate ranked images by applying different levels of distortions to an image. As an example, given a reference image we can apply various levels of Gaussian blur. The set of images which is thus generated can be easily ranked because we do know that adding Gaussian blur (or any other distortion) always deteriorates the quality score. Note that in such a set of ranked images we do not have any absolute IQA scores for any images – but we do know for any pair of images which is of higher quality. See Fig. 2 (bottom) for an illustration of this.
For the crowd counting problem, we can obtain ranked sets of images by comparing parts of the same image which are contained within each other: an image which is contained by another image will contain the same number or fewer persons than the larger image. This fact can be used to generate a large dataset of ranked images from unlabeled crowd images. See Fig. 4 (bottom) for an illustration of this process for crowd counting.
In the following sections we will show that ranked data can be used to train networks for regression problems. We are especially interested in domains where the ranked data can be automatically generated from images of the problem domain without requiring any additional human labeling. This allows us to create large dataset of ranked data. We consider regression to be the principal task of the network, and we refer to the ranking task as a self-supervised proxy task. It is self-supervised since the ranking task is an additional task for which data is freely available and no annotation is required.
3.2 Multi-task regression and ranking
In this section, we formalize the problem of training from both labeled and ranked data for regression problems. We consider a regression problem where we have a dataset of observed data in pairs:
where are images and . The input images and target variables are assumed to be related by some unknown function . The aim is to find a function with parameters that captures (and generalizes) the relationship between input and output . The parameters of the regression function are usually fit by minimizing an empirical risk over training examples , for example the squared Euclidean loss:
In our formulation, is a deep Convolutional Neural Network (CNN), and minimizing is done with stochastic gradient descent (SGD). We refer to the regression task as the principal task, since the final objective is to accurately estimate such a regression.
We also assume that we have a function which we can apply to input images. These functions are special in that the parameter space is ordered and that for any parameters and any image :
What this means is that we have a partial order in the parameter space that induces an ordering in the proxy task space. An example of a possible function would be adding an image distortion parameterized by a single scalar number: increasing this parameter in implies decreasing image quality in . For crowd counting we can parameterize using rectangular crops: if a rectangle is entirely contained in rectangle , then the number of persons in is less than or equal to . In other words: .
We can now apply this function to generate an auxiliary dataset consisting of ranked images. Importantly, the ordering of the parameter space and the application of is independent of any labeling of training data. The dataset could contain images which are present in dataset but in addition it could also contain images not in and for which we have no annotations. We can use any input data relevant to the domain in order to generate the dataset of ranked images. Since respects this ranking condition under application of functions , we can use the functions to generate training data for learning .
From dataset we can train a network by minimizing the ranking hinge loss according to:
where is a margin, and are (potentially unlabeled) images transformed into ranked images after applying and for . Training a network with Eq. (4) yields a network which can rank images, and we will refer to the task of ranking as the self-supervised proxy task.
The most common approach to minimizing losses like uses a Siamese network 
, which is a network with two identical branches connected to a loss module. The two branches share weights during training. Pairs of images and labels are the input of the network, yielding two outputs which are passed to the loss. The gradients of the loss function with respect to all model parameters are computed by backpropagation and updated by the stochastic gradient method (e.g. SGD). For problems where both labeled data (like in dataset) and ranked data (like in dataset ) are present, we can optimize the network using both sources of data using a multi-task loss:
where is a tradeoff parameter that balances the relative weight of the losses. It is important to note here that we consider the same function for the regression and the ranking loss. In practice this means that the three networks, one for regression, and the two networks used in the Siamese network, have the same architecture and share their parameters.
One could also consider different ways to combine both the labeled dataset and the ranking dataset. In earlier work we investigated using the ranking dataset to train initial weights of the network, after which we used the labeled data for fine-tuning . This is the approach which is used by almost all self-supervised methods in computer vision [1, 4, 3, 5, 15]. However, in  we found this to be inferior to using the multi-task loss, and in this work only consider multi-task formulations like in Eq. (5).
3.3 Efficient Siamese backpropagation
One drawback of Siamese networks is redundant computation. Consider all possible image pairs constructed from three images. In a standard implementation all three images are passed twice through the network, because they each appear in two pairs. Since both branches of the Siamese network are identical, we are essentially doing twice the work necessary since any image need only be passed once through the network. It is exactly this idea that we exploit to render backpropagation more efficient for Siamese network training. In fact, nothing prevents us from considering all possible pairs in a mini-batch, with hardly any additional computation. We add a new layer to the network that generates all possible pairs in a mini-batch at the end of the network right before computing the loss. This eliminates the problem of pair selection and boosts efficiency. At the end of this section we discuss how our approach compares to works [12, 13] published concurrently with ours and observed similar efficiency gains.
To appreciate the speed-up of efficient Siamese backpropagation consider the following. If we have one reference image distorted from which we have generated ranked images using , then for a traditional implementation of the Siamese network we would have to pass a total of images through the network – which is twice the number of pairs you can generate with images. Instead we propose to pass all images only once and consider all possible pairs only in the loss computation layer. This reduces computation to just images passed through the network. Therefore, in this case the speed-up is equal to: . In the best scenario is equal to the number of images in the mini-batch, and hence the speed-up of this method would be in the order of the mini-batch size. Due to the high correlation among the set of all pairs in a mini-batch, we expect the final speedup in convergence to be lower.
To simplify notation in the following, assume we have an image and two transformation parameters . Letting , the contribution to the ranking loss of these two images can be written as:
The gradient of this term from with respect to the model parameters is:
This gradient of above is a sum since the model parameters are shared between both branches of the Siamese network and and are computed using exactly the same parameters.
Considering all pairs in a mini-batch of size , the loss from Eq. (4) can then be written as:
The gradient of the mini-batch loss with respect to parameter can then be written as:
We can now express the gradient of the loss function of the mini-batch in matrix form as:
is the vector of all ones of length. For a standard single-branch network, we would average the gradients for all batch samples to obtain the gradient of the mini-batch. This is equivalent to setting
to the identity matrix in Eq. (10) above. For Siamese networks where we consider all pairs in the mini-batch we obtain Eq. (9) by setting to:
For the ranking hinge loss we can write:
and is used to indicate the ordering of the parameters used to generate the images to which was applied to derive outputs and :
The above analysis works for different parameter settings on the same source image. When considering multiple source images in a mini-batch, only different parameter settings on the same image are considered comparable when defining .
Generally, the complexity of training Siamese networks is ameliorated via different pair sampling techniques. In , the authors propose a hard positive and hard negative mining strategy to forward propagate a set of pairs and sample the highest-loss pairs for backpropagation. However, hard mining comes with a high computational cost (they report an increase of up to 80% of total computation cost). In  the authors propose semi-hard pair selection, arguing that selecting hardest pairs can lead to bad local minima. In  the authors take a batch of pairs as input and choose the four hardest negative samples within the minibatch. In parallel with our work on fast backpropagation for Siamese networks [14, 15] several similar methods have been developed [12, 13]. To solve for bad local optima,  optimize a smooth upper bound loss function. This is implemented by considering all possible pairs in a mini-batch after forwarding the images through the network. In , the -pair Loss is proposed to compute pairwise similarity within the batch to construct negative examples instead of one in triplet loss. Hard negative class mining is applied to improve convergence speed. Both these works, like ours, prevent the redundant computation which is introduced by the multiple branches in the Siamese network.
3.4 Active learning from rankings
The objective of active learning is to reduce the cost of labeling by prioritizing the most informative samples for labeling first. Rather than labeling randomly selected examples from a pool of unlabeled data, active learning methods analyze unlabeled data with the goal of identifying images considered difficult and that therefore are more valuable if labeled. Especially for deep networks, which require many examples, as well as for applications for which labeling is very costly, active learning is an important and active area of research.
We show here how the self-supervised proxy task can be leveraged for active learning. For this purpose we define a function which estimates the certainty of the current network on a specific image . This estimate allows us to order the available unlabeled dataset according to certainty. Labeling the images for which the algorithm is uncertain is then expected to yield larger improvement in performance of the network than just randomly adding images.
The certainty function is defined as:
where if predicate is true or false otherwise, and is the number of sampled parameter pairs applied to each image . As described in the previous section, the pairs of parameters can be used to generate ranked images via the function . Again, these pairs can be automatically computed and no annotation is required.
By design, we know that . Given an unlabeled dataset and a current state of the trained network , we perform the proxy task for a total of times on each image and then compute . We then label images starting from low confidence to high confidence. The process is detailed in Algorithm 1.
4 Image Quality Assessment by Learning to Rank
In this section we apply the framework proposed in the previous section to the problem of Image Quality Assessment (IQA) . IQA aims to automatically predict the perceptual quality of images. IQA estimates should be highly correlated with quality assessments made by a range of very many human evaluators (commonly referred to as the Mean Opinion Score (MOS) [53, 54]). We focus here on no-reference IQA (NR-IQA), which refers to the case where the undistorted image (called reference image) is not available during the quality assessment.
: our network for no-reference IQA uses a VGG16 network pretrained on ImageNet. We decapitate the network and replace the original head with a new fully-connected layer generating a single output.Bottom: pairs with known ranking are generated by distorting images with standard, parametric distortions. Increasing the distortion level guarantees that the images are of progressively worse quality.
The application of CNNs to IQA has resulted in significant improvements compared to previous hand-crafted approaches [6, 7, 8]. These methods had to train their networks on the small available datasets for IQA. Further improvements would be expected if larger datasets were made available. However, the annotation of IQA images is a labor intensive task, which requires multiple human annotators for every image. It is therefore an interesting application field to evaluate our framework, since by adding ranking as a proxy task, we are able to add unlabeled data during the training process.
We first discuss existing datasets for IQA and how to automatically generate IQA rankings. Then we introduce the network which we train for the IQA task, together with some application-specific training choices.
4.1 IQA datasets
We perform experiments on two standard IQA datasets:
LIVE : Consists of 808 images generated from 29 original images by distorting them with five types of distortion: Gaussian blur (GB), Gaussian noise (GN), JPEG compression (JPEG), JPEG2000 compression (JP2K) and fast fading (FF). The ground-truth Mean Opinion Score for each image is in the range [0, 100] and is estimated using annotations by 161 human annotators.
TID2013 : Consists of 25 reference images with 3000 distorted images from 24 different distortion types at 5 degradation levels. Mean Opinion Scores are in the range [0, 9]. Distortion types include a range of noise, compression, and transmission artifacts. See the original publication for the list of specific distortion types.
4.2 Generating ranked image sets for IQA
The IQA datasets LIVE and TID2013 are derived from only 29 and 25 original images, respectively. Deep networks trained on so few original images will almost certainly have difficulty generalizing to other images. Here we show how to automatically generate rankings from arbitrary images which can be added during the training process (i.e. we show what function in Eq. (4) we use to generate IQA rankings).
Using an arbitrary set of images, we can synthetically generate deformations of these over a range of distortion intensities. As an example, consider Fig. 2 (bottom) where we have distorted an image with increasing levels of Gaussian noise. Although we do not know the absolute IQA score, we do know that images which are more distorted should have a lower score than images which are less distorted. This fact can be used to generate huge datasets of ranked images. We can use a large variety of images and distortions to construct the datasets of ranked images.
We use the Waterloo  dataset, which consists of 4,744 high quality natural images carefully chosen from the Internet, to generate a large ranking dataset. To test on the LIVE database, we generate four types of distortions which are widely used and common: Gaussian Blur (GB), Gaussian Noise (GN), JPEG, and JP2K. We apply these distortions at five levels, resulting in a total of 20 distortions for all images in the Waterloo dataset. To test on TID2013, we generate 17 out of a total of 24 distortions at five levels, yielding a total of 85 distortions per image (see the Appendix for more details). There were seven distortions which we could not generate, however we found that adding the other distortions also resulted into improvements for the distortions for which we could not generate rankings.
4.3 IQA network
For the IQA problem we have a training dataset with images and ground truth image quality . After generating the ranked image dataset we can train an IQA network. As a CNN backbone we use the VGG-16 network, only changing the last layer to output a single IQA score. With respect to the basic architecture given in Fig. 1, here we specify the architecture we use for IQA estimation as shown in Fig. 2 (top). For the labeled data we use the Euclidean distance between the prediction of the network and the ground truth as the regression loss:
and will optimize this loss jointly with the ranking loss as proposed in Eq. (5).
We randomly sample sub-images from the original high resolution images. We do this instead of scaling to avoid introducing distortions caused by interpolation or filtering. The size of sampled images is determined by each network. However, the large size of the input images is important since input sub-images should be at least 1/3 of the original images in order to capture context information. This is a serious limitation of the patch sampling approach[7, 8] that samples very small patches from the original images. In our experiments, we sample pixel images from original images varying from to pixels.
5 Crowd Counting by Learning to Rank
In this section, we adapt the proposed framework to the task of crowd counting. Perspective distortion, clutter, occlusion, non-uniform distribution of people, complex illumination, scale variation, and a host of other scene-incidental imaging conditions render person counting and crowd density estimation in unconstrained images an daunting problem.
Techniques for crowd counting have been recently improved using Convolutional Neural Networks (CNNs). As with most CNN architectures, however, these person counting and crowd density estimation techniques are highly data-driven. For person counting, the labeling burden is even more onerous than usual. Training data for person counting requires that each individual person be meticulously labeled in training images. It is for this reason that person counting and crowd density estimation datasets tend to have only a few hundred images available for training. As a consequence, crowd counting networks are expected to benefit from additional unlabeled data which is used to train a ranking proxy task.
5.1 Crowd counting datasets
We use two standard benchmark crowd counting datasets:
UCF_CC_50 : This dataset contains 50 annotated images of different resolutions, illuminations and scenes. The variation of densities is very large among images from 94 to 4543 persons with an average of 1280 persons per image.
ShanghaiTech : Consists of 1198 images with 330,165 annotated heads. This dataset includes two parts: 482 images in Part_A which are randomly crawled from the Internet, and 716 images in Part_B which are taken from busy streets. Both parts are further divided into training and evaluation sets. The training and test of Part_A has 300 and 182 images, respectively, whereas that of Part_B has 400 and 316 images, respectively.
Ground truth annotations for crowd counting typically consist of a set of coordinates which indicate the ’center’ (typically head center of a person). To convert this data to crowd density maps we place a Gaussian with standard deviation of 15 pixels and sum these for all persons in the scene to obtain. This is a standard procedure and is also used in [9, 10].
5.2 Generating ranked image sets for counting
Here we show how to automatically generate the rankings from unlabeled crowd counting images. The main idea is based on the observation that all patches contained within a larger patch must have a fewer or equal number of persons than the larger one (see Fig. 4). This observation allows us to collect large datasets of crowd images with known relative ranks. Rather than having to painstakingly annotate each person we are only required to verify if the image contains a crowd. Given a crowd image we extract ranked patches according to Algorithm 2.
convolutional layer with stride 1 which should estimate local crowd density. A sum pooling layer is added to the ranking channel of the network to arrive at a scalar value whose relative rank is known.Bottom: image pairs with known relative ranks are generated by selectively cropping unlabeled crowd images so that successive crops are entirely contained in previous ones.
To collect a large dataset of crowd images from the Internet we use two different approaches:
Keyword query: We collect a crowd scene dataset from Google Images by using different key words: Crowded, Demonstration, Train station, Mall, Studio, Beach, all of which have high likelihood of containing a crowd scene. Then we delete images not relevant to our problem by simple visual inspection. In the end, we collected 1,180 high resolution crowd scene images, which is about 24x the size of the UCF_CC_50 dataset, 2.5x the size of ShanghaiTech Part_A, and 2x the size of ShanghaiTech Part_B. Note that no other annotation of images is performed. Example images from this dataset are given in Fig. 3 (top row).
Query-by-example image retrieval:For each annotated benchmark dataset, we collect an unlabeled dataset using the training images as queries with the Google Images visual image search engine. We choose the first ten similar images and remove irrelevant ones. For UCF_CC_50 we collected 256 images, for ShanghaiTech Part_A 2229 images, and for ShanghaiTech Part_B 3819 images. An example of images returned for a specific query image is given in Fig. 3 (bottom row).
|Input:||A crowd scene image, number of patches and scale factor .|
|Step 1:||Choose an anchor point randomly from the anchor region. The anchor region is defined to be the size of the original image, centered at the original image center, and with the same aspect ratio as the original image.|
|Step 2:||Find the largest square patch centered at the anchor point and contained within the image boundaries.|
|Step 3:||Crop additional square patches, reducing size iteratively by a scale factor . Keep all patches centered at anchor point.|
|Step 4:||Resize all patches to input size of network.|
|Output:||A list of patches ordered according to the number of persons in the patch.|
5.3 Crowd density estimation network
We first explain the network architecture which is trained on available crowd counting datasets with ground truth annotations (see Fig. 4) . This network regresses to a crowd density image which indicates the number of persons per pixel (examples of such maps are given in Fig. 6). A summation of all values in such a crowd density image gives an estimate of the number of people in the scene. In the experimental section we consider this network as the baseline method to which we compare.
Our baseline network is derived from VGG-16 
pre-trained on ImageNet. VGG-16 consists of 13 convolutional layers followed by three fully connected layers. We adapt the network to regress to person density maps by removing its three fully connected layers and the last max-pooling layer (pool5) to prevent further reduction of spatial resolution. In their place we add a single convolutional layer (a
filter with stride 1 and zero padding to maintain same size) which directly regresses to the crowd density map. As the counting losswe use the Euclidean distance between the estimated and ground truth density maps, as given by:
where is the number of images in a training batch, is ground truth person density map of the -th image in the batch, and the prediction from the network as .
To further improve the performance of our baseline network, we introduce multi-scale sampling from the available labeled datasets during training. Instead of using the whole image as an input, we randomly sample square patches of varying size (from 56 to 448 pixels). In the experimental section we verify that this multi-scale sampling is important for good performance. Since we are processing patches rather than images we use to refer to the estimate of patch from now on. The importance of multi-scale processing of crowd data was also noted in .
Finally, we add a summation layer to the network. This summation layer takes as an input the estimated density map and sums it to a single number (the estimate of the number of persons in the image). This output is used to compute the ranking loss (see Eq. (4)) for the unlabeled images in the ranked dataset. With respect to the basic architecture in Fig. 1, we use a sum polling layer as a ranking specific layer as shown in Fig. 4 (top).
6 Experimental results
In this section we report on an extensive set of experiments performed to evaluate the effectiveness of learning from rankings for Image Quality Assessment in section 6.1 and crowd counting in section 6.2
. We use the Caffe deep learning framework in all the experiments.
6.1 Image Quality Assessment (IQA)
We performed a range of experiments designed to evaluate the performance of our approach with respect to baselines and the state-of-the-art in IQA. These experiments make use of standard datasets (for benchmark evaluation) and an additional dataset used for generating ranked distorted image pairs as explained in section 4.
We use Stochastic Gradient Descent (SGD) with an initial learning rate of 1e-4 for efficient Siamese network training and 1e-6 for fine-tuning. Training rates are decreased by a factor of 0.1 every 10,000 iterations for a total of 50,000 iterations. For both training phases we use weight decay (weight 5e-4). For multi-task training, we use in Eq. (5) with a learning rate of 1e-6 on the LIVE dataset and 1e-5 on TID2013. We found to work well in initial experiments, but cross validating
on held-out data is expected to improve results for specific datasets. During training we sample a single subimage from each training image per epoch. When testing, we randomly sample 30 sub-images from the original images, as suggested in, and pass all the trained models. The average of all outputs of the sub-regions is the final score for each distorted image.
Two evaluation metrics are traditionally used to evaluate the performance of IQA algorithms: the Linear Correlation Coefficient (LCC) and the Spearman Rank Order Correlation Coefficient (SROCC). LCC is a measure of the linear correlation between the ground truth and the predicted quality scores. Givendistorted images, the ground truth of -th image is denoted by , and the predicted score from the network is . The LCC is computed as:
where and are the means of the ground truth and predicted quality scores, respectively.
Given distorted images, the SROCC is computed as:
where is the rank of the ground-truth IQA score in the ground-truth scores, and is the rank of in the output scores for all images. The SROCC measures the monotonic relationship between ground-truth and estimated IQA.
6.1.1 Ablation study
In this experiment, we evaluate the effectiveness of using rankings to estimate image quality. We compare our multi-task approach with different baselines: fine-tuning the VGG-16 network initialized from ImageNet to obtain the mapping from images to their predicted scores (which we call Baseline in our experiments), and two other baselines from our previous work : VGG-16 (initialized pre-trained ImageNet weights) trained on ranking data (called RankIQA), and our RankIQA approach fine-tuned on TID2013 after training using ranked pairs of images (called RankIQA+FT). In order to evaluate the effectiveness of our sampling method in the multi-task setting, we also report the accuracy for multi-task training (called MT-RankIQA) with different sampling methods: standard random pair sampling, and a hard-negative mining method similar to . For standard random pair sampling we randomly choose 36 pairs for each mini-batch from the training sets. For the hard negative mining strategy we start from 36 pairs in a mini-batch, and gradually increase the number of hard pairs every 5000 iterations. For our method we pass 72 images in each mini-batch. With these settings the computational costs for all three methods are equal, since at each iteration 72 images are passed through the network.
We follow the experimental protocol used in HOSA . The entire TID2013 database including all types of distortions is divided into 80% training images and 20% testing images according to the reference images and their distorted versions. The results are shown in Table I, where ALL means testing all distortions together. All the experiments are performed 10 times and the average SROCC is reported
From Table I, we can draw several conclusions. First, it is hard to obtain good results by training a deep network directly on IQA data. This is seen in the Baseline results and is due to the scarcity of training data. Second, competitive results are obtained using RankIQA without access to the ground truth of IQA dataset during training the ranking network, which strongly demonstrates the effectiveness of training on ranking data. The RankIQA-trained network alone does not provide accurate IQA scores (since it has never seen any) but does yield high correlation with the IQA scores as measured by SROCC. After fine-tuning on the TID2013 database, we considerably improve performance for all sampling methods: with random sampling we improve on the baseline by 11%, while our efficient sampling method further outperforms random sampling by 2.4% in terms of LCC (similar conclusions can be drawn from the SROCC results.
Finally, in Table I we also compare the three optimization methods for multi-task training: random pair sampling, hard negative mining, and our proposed efficient Siamese backpropagation. We see that our efficient back-propagation strategy with multi-task setting obtains the best results, further improving the accuracy by 2.8% on LCC and 2.6% on SROCC with respect to RankIQA+FT. This clearly shows the benefits of the proposed backpropagation scheme. In the next section, we use MT-RankIQA to refer to our method trained with the multi-task loss using our efficient Siamese backpropagation method.
To demonstrate the ability of our approach to generalize to unseen distortions, we trained our multi-task approach with different numbers of synthetic distortions as auxiliary data combined with all labeled data in the TID2013 dataset. All results are the average of three runs. As shown in Table III, adding more distortions results in increased overall performance on the test set. Note also that adding specific distortions not only consistently improves accuracy on seen distortions consistently, but also on unseen ones.
|Overall LCC accuracy||Average gain|
|Seen distortions||Unseen distortions|
6.1.2 Comparison with the state-of-the-art
We compare the performance of our method with state-of-the-art Full-reference IQA (FR-IQA) and NR-IQA methods on both TID2013 and LIVE dataset.
Evaluation on TID2013. Table II includes results of state-of-the-art methods. We see that for several very challenging distortions (14 to 18), where all other methods fail, we obtain satisfactory results. For individual distortions, there is a huge gap between RankIQA and other methods on most distortions. The state-of-the-art method HOSA performs slightly better than our methods on 6 out of 24 distortions. For all distortions, RankIQA+FT achieves about 5% higher than HOSA, and about 3% more is gained by using multi-task training. Our methods also perform well on distortions for which were unable to generate rankings. This indicates that different distortions share some common representation and training the network jointly on all distortions also improves results for the distortions for which we did not generate rankings.
Evaluation on LIVE. As done in [7, 68], we randomly split the reference images and their distorted version from LIVE into 80% training and 20% testing sample and compute the average LCC and SROCC scores on the testing set after training to convergence. This process is repeated ten times and the results are averaged. These results are shown in Table IV. For fair comparison with the state-of-the-art, we train our ranking model on four distortions (all but FF), but we fine-tune our model on all five distortions in the LIVE dataset to compute ALL. As shown in Table IV our approach improves by 0.4% and 1% in LCC and SROCC, respectively, the best NR-IQA results reported on ALL distortions. This indicates that our method outperforms existing work including the current state-of-the-art NR-IQA method SOM  and DNN , and even achieves competitive results as state-of-the-art FR-IQA method DCNN  which, being a full-reference approach, has the benefit of having access to the high-quality original reference image.
6.1.3 Active learning for IQA
We demonstrate the effectiveness of active learning on the IQA problem (see Algorithm 1). As a dataset we consider all 24 distortions for each of 19 reference images from TID2013, yielding a dataset of 456 image-distortion pairs. Each image is distorted for each distortion at five distortion levels. During active learning we aim to select which images with what particular distortion are expected to most improve performance. We add all five distortion levels of the selected image with the particular distortion to the labeled pool. The active learning loop starts with 10% of this data labeled; hence the labeled samples is 10% and the examples consist of the remaining 90% of the data without labels. We then perform active learning cycles. In each cycle images with a particular distortion are added incrementally, using the full training set from the previous iteration to estimate informativeness for the remaining unlabeled data. We use in the experiment, and compare results to a baseline of randomly adding 10% additional labeled samples at each step.
Results for active learning are given in Fig. 5. It is clear that our active learning algorithm obtains results superior to random selection. When adding an additional 10% of data (using a total of 20% labeled data) we achieve similar accuracy as adding an additional 40% data (using a total of 50% labeled data) with random selection, thereby reducing the labeling cost by 75% compared to the baseline.
6.2 Crowd counting
Here we report on a range of experiments evaluating our approach with respect to baselines and the state-of-the-art methods for crowd counting.
We use SGD with a batch size of 25 for both ranking and counting, and thus a batch size of 50 for multi-task training. For the ranking plus fine-tuning method, the learning rate is 1e-6 for both ranking and fine-tuning. For multi-task training, we found that yielded good results on all datasets. Cross validating on held-out data is expected to improve results for specific datasets. Learning rates are decreased by a factor of 0.1 every 5,000 iterations for a total of 10K iterations. For both training phases we use weight decay with a weight of 5e-4. During training we sample one sub-image from each training image per epoch. We perform down-sampling of three scales and up-sampling of one scale on the UCF_CC_50 dataset and only up-sampling of one scale on the ShanghaiTech dataset. The number of ranked crops , the scale factor , and the anchor region (see Algorithm 2).
Following existing work, we use the mean absolute error (MAE) and the mean squared error (MSE) to evaluate different methods. These are defined as follows:
where is the number of test images, is the ground truth number of persons in the th image, and is number of persons predicted by the network in the th image.
|Method||Split 1||Split 2||Split 3||Split 4||Split 5||Ave MAE|
|+ Pre-trained model||570.01||350.63||334.89||184.79||202.41||328.54|
6.2.1 Ablation study.
We begin with an ablation study on the UCF_CC_50 dataset. The aim is to evaluate the relative gain of the proposed improvements and to evaluate the use of a ranking loss against the baseline. The ranked images in this experiment are generated from the Keyword dataset. The results are summarized in Table V. We can immediately observe the benefit of using a pre-trained ImageNet model in crowd counting, with a significant drop in MAE of around 28% compared to the model trained from scratch. By using both multi-scale data augmentation and starting from a pre-trained model, another improvement of around 6% is obtained.
The Ranking+FT method performs worse than directly fine-tuning from a pre-trained ImageNet model. This is probably caused by the poorly-defined nature of the self-supervised task. To optimize this task the network could decide to count anything, e.g. ‘hats’, ‘trees’, or ‘people with red shirts’, or even just ‘edges’ – all of which would satisfy the ranking constraints that are imposed.
Next, we compare the three sampling strategies for combining the ranking and counting losses for multi-task training. When using multi-task training with random pair sampling, the average MAE is reduced by about 15 points. Hard mining obtains about 5 points average MAE less than random sampling. However, our efficient back-propagation approach reduces the MAE further to 279.6. This shows that by jointly learning both the self-supervised and crowd counting tasks, the self-supervised task is forced to focus on counting persons. Given its superior results, we consider only the “Multi-task” with efficient back-propagation approach for the remainder of the experiments.
|Idrees et al. ||419.5||541.6|
|Onoro et al. ||333.7||425.2|
|Walach et al. ||364.4||341.4|
6.2.2 Comparison with the state-of-the-art
Evaluation on the UCF_CC_50 dataset. A five-fold cross-validation was performed for evaluating the methods. Results are shown in Table VI. Our multi-task training method using the unlabeled Keyword Dataset reduces the MAE from 291.0 to 279.6, which is comparable to ACSCP  which was published at the same time as our original work. Our approach performs slightly worse than the state-of-the-art methods CSRNet  and ic-CNN  (also published around the same time as our original work), but in general our model has fewer parameters than CSRNet and a simpler inference procedure compared to ic-CNN. However, the MSE of our method on UCF_CC_50 dataset is worse than the state-of-the-art methods [69, 36, 44], but achieves competitive results compared to [41, 39]. This indicates that our method and also 
work better in general but have more extreme outliers. Compared to training on the Keyword dataset, learning from the Query-by-example dataset is slightly worse, which might be because most images from UCF_CC_50 are black and white with low resolution, which often does not lead to satisfactory query results. An example of prediction in UCF_CC_50 using our network is shown in Fig.6.
|Part A||Part B|
|Ours: Multi-task (Query-by-example)||72.0||106.6||14.4||23.8|
|Ours: Multi-task (Keyword)||73.6||112.0||13.7||21.4|
Evaluation on the ShanghaiTech dataset. Looking at Table VII, we can draw conclusions similar to those on UCF_CC_50. We see here that using the Query-by-example Dataset further improves by about 2% on ShanghaiTech – especially for Part_A, where our approach surpasses the state-of-the-art method , but is still slightly worse than CSRNet  and ic-CNN . An example of prediction by our network on ShanghaiTech is given in Fig. 6. For comparison, we also provide the results of our baseline method (including fine-tuning from a pre-trained model and multi-scale data augmentation) on this dataset: and on Part A, and and on Part B.
Evaluation on the WorldExpo’10 dataset The WorldExpo’10 dataset  consists of 3980 frames of size 576 720 from 1132 video sequences captured by 108 surveillance cameras. The dataset is split into training set with 103 scenes and test set consisting of 5 different scenes. Regions of interest (ROIs) are provided for the whole dataset, which are used as a mask during testing. We consider training directly on the only ground truth labels as a baseline, and for our multi-task approach, when we test on one specific scene, the rest of scenes in the test set are used to generate the ranked image set. We compare our multi-task training to the baseline and other state-of-the-art methods in Table VIII. It is clear that our multi-task training approach outperforms the baseline in all five cases and achieves comparable results compared to other methods in terms of MAE.
|Method||Scene 1||Scene 2||Scene 3||Scene 4||Scene 5||Ave MAE|
Evaluation on the UCSD dataset The UCSD dataset  has 2000 frames with a region of interest (ROI) varying from 11 to 46 persons per image. The resolution of each frame is fixed and small (238 158), so we change the input of network to 112 112 by removing all layers after pool4 in VGG-16. Thus the output retains the same 1/16 of input size. We follow the same settings as , using frames between 600 and 1400 as training set, and the rest as test set. In order to train our multi-task approach and compare fairly to other methods, we generate ranked sets using the frames from 1 to 600 and test on the frames from 1401 to 2000 (and vice versa). We do not collect additional unlabeled data since the training and test set are from the same camera. Results are shown in Table IX. Our baseline performs similarly to the state-of-the-art methods [11, 39], but slightly worse than ACSCP  and MCNN . Our multi-task approach reduces the baseline MAE from 1.60 to 1.17. Our multi-task approach works on less dense datasets like UCSD because, though the baseline might make incorrect predictions on patches with the almost same number of headcounts, our learning-to-rank branch can constrain it and ensure correct ranking predictions.
Evaluation on the UCF-QNRF dataset. UCF-QNRF  is a very challenging dataset consisting of 1,535 images with average of 815 people per image. The average resolution of images is much larger compared to other datasets, with images up to 6,0009,000 pixels. We resize all images to have a maximum dimension of 1024 pixels without changing the aspect ratio. The Keyword Dataset is used as unlabeled data to for the multi-task learning. Results are shown in Table X, which indicate that our technique consistently improves counting performance – even on datasets like UCF-QNRF with significantly more labeled training samples. Compared to our baseline, the MAE is reduced from 137 to 124 and MSE is down to 196. Our method outperforms all other methods including CompositionLoss  in MAE and performs slightly worse in MSE.
Evaluation on transfer learning. As proposed in 
, to demonstrate the generalization of the learned model, we test our method in the transfer learning setting by using Part_A of the ShanghaiTech dataset as the source domain and using UCF_CC_50 dataset as the target domain. The model trained on Part_A of ShanghaiTech is used to predict the crowd scene images from UCF_CC_50 dataset, and the results can be seen in TableXI. Using only counting information improves the MAE by 12% compared to reported results in . By combining both ranking and counting datasets, the MAE decreases from 349.5 to 337.6, and MSE decreases from 475.7 to 434.3. In conclusion, these results show that our method significantly outperforms the only other work reporting results on the task of cross-dataset crowd counting.
6.2.3 Active learning for crowd counting
We use the Shanghai Part_A dataset to evaluate our active learning approach on Crowd counting. We use of the training set as the initially labeled training samples (and the remaining 90%), and set to add 10% of the training images in each of the active learning cycles (see Algorithm 1). We use in this experiment to evaluate ranking certainty. Again we compare to the baseline of randomly selecting images from . The result is shown in Fig. 7. Our approach performs consistently better than random selection in terms of MAE. We obtain similar results with 20% of the training data as random approach does with 40% – reducing the labeling effort by 50%. These results also show that performance saturates after 60% and no further improvement is obtained by adding the last 40% images considered least useful by the active learning algorithm. The results clearly show that our active learning method correctly identifies the images, which when labeled, contribute most to an improved crowd counting network.
In this article we explored ranking as a self-supervised proxy task for regression problems. For many regression problems the collection of supervised data is an expensive and laborious process. We showed, however, that there exist some problems for which it is easy to obtain ranked image sets, and that these ranked image sets can be exploited to improve the training of the network. In addition, we proposed a method for fast backpropagation for the ranking loss. This method removes the redundant computation which is introduced by the multiple branches of Siamese networks, and instead uses a single branch after which all possible pairs of the minibatch are combined (rather than just a selection of pairs).
We applied the proposed framework to two regression problems: Image Quality Assessment and crowd counting. In the case of Image Quality Assessment, the ranked data sets are formed by adding increasing levels of distortions to images. For crowd counting, ranked sets are formed by comparing image crops which are contained within each other: a smaller image contained in another larger one will contain the same number or fewer persons than the larger image. Experimental results show that for both applications results improve significantly when adding unlabeled data for the ranking task. In addition, we have shown that the best results are obtained when using our efficient backpropagation method in a multi-task setting.
We have also shown that the proxy task can be used as an informativeness measure for unlabeled images. The number of errors made on the proxy task can be used to drive an active learning algorithm to select the best images to label from a pool of unlabeled ones. These images, once added to the training set, will most improve the performance of the network. Experimental results show that, for both IQA and crowd counting, this method can reduce the labeling effort by a large margin.
We acknowledge the Spanish project TIN2016-79717-R, the CHISTERA project M2CR (PCIN-2015-251) and the CERCA Programme / Generalitat de Catalunya. Xialei Liu acknowledges the Chinese Scholarship Council (CSC) grant No.201506290018. We also acknowledge the generous GPU donation from NVIDIA.
-  C. Doersch, A. Gupta, and A. A. Efros, “Unsupervised visual representation learning by context prediction,” in ICCV, 2015.
G. Larsson, M. Maire, and G. Shakhnarovich, “Colorization as a proxy task for visual understanding,” inCVPR, 2017.
-  R. Zhang, P. Isola, and A. A. Efros, “Colorful image colorization,” in ECCV, 2016.
-  D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros, “Context encoders: Feature learning by inpainting,” in CVPR, 2016.
-  M. Noroozi, H. Pirsiavash, and P. Favaro, “Representation learning by learning to count,” in ICCV, 2017.
-  Y. Liang, J. Wang, X. Wan, Y. Gong, and N. Zheng, “Image quality assessment using similar scene as reference,” in ECCV. Springer, 2016, pp. 3–18.
-  L. Kang, P. Ye, Y. Li, and D. Doermann, “Convolutional neural networks for no-reference image quality assessment,” in CVPR, 2014, pp. 1733–1740.
-  L. Kang, P. Ye, Y. Li, and D. Doermann, “Simultaneous estimation of image quality and distortion via multi-task convolutional neural networks,” in ICIP. IEEE, 2015, pp. 2791–2795.
-  D. Onoro-Rubio and R. J. López-Sastre, “Towards perspective-free object counting with deep learning,” in ECCV. Springer, 2016.
-  Y. Zhang, D. Zhou, S. Chen, S. Gao, and Y. Ma, “Single-image crowd counting via multi-column convolutional neural network,” in CVPR, 2016.
-  D. Babu Sam, S. Surya, and R. Venkatesh Babu, “Switching convolutional neural network for crowd counting,” in CVPR, 2017.
-  H. O. Song, Y. Xiang, S. Jegelka, and S. Savarese, “Deep metric learning via lifted structured feature embedding,” in CVPR, 2016.
-  K. Sohn, “Improved deep metric learning with multi-class n-pair loss objective,” in NIPS, 2016, pp. 1857–1865.
-  X. Liu, “Learning from rankings for no-reference image quality assessment by siamese network,” Master thesis of the Universitat Autonoma de Barcelona, 2016.
-  X. Liu, J. van de Weijer, and A. D. Bagdanov, “Rankiqa: Learning from rankings for no-reference image quality assessment,” in ICCV, 2017.
-  X. Liu, J. van de Weijer, and A. D. Bagdanov, “Leveraging unlabeled data for crowd counting by learning to rank,” in CVPR, 2018.
-  W. Chen, T.-Y. Liu, Y. Lan, Z.-M. Ma, and H. Li, “Ranking measures and loss functions in learning to rank,” in NIPS, 2009.
-  D. Sculley, “Large scale learning to rank,” in NIPS Workshop on Advances in Ranking, 2009, pp. 1–6.
B. Settles, “Active learning,”
Synthesis Lectures on Artificial Intelligence and Machine Learning, vol. 6, no. 1, pp. 1–114, 2012.
-  Z. Zhou, J. Shin, L. Zhang, S. Gurudu, M. Gotway, and J. Liang, “Fine-tuning convolutional neural networks for biomedical image analysis: actively and incrementally,” in CVPR, 2017, pp. 7340–7349.
-  D. P. Papadopoulos, J. R. Uijlings, F. Keller, and V. Ferrari, “We don’t need no bounding-boxes: Training object class detectors using only human verification,” in CVPR, 2016.
-  L. Yang, Y. Zhang, J. Chen, S. Zhang, and D. Z. Chen, “Suggestive annotation: A deep active learning framework for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2017, pp. 399–407.
-  D. D. Lewis and J. Catlett, “Heterogeneous uncertainty sampling for supervised learning,” in Machine Learning Proceedings 1994. Elsevier, 1994, pp. 148–156.
T. Scheffer, C. Decomain, and S. Wrobel, “Active hidden markov models for information extraction,” inInternational Symposium on Intelligent Data Analysis. Springer, 2001, pp. 309–318.
-  B. Settles, M. Craven, and S. Ray, “Multiple-instance active learning,” in NIPS, 2008, pp. 1289–1296.
-  B. Settles, M. Craven, and L. Friedland, “Active learning with real annotation costs,” in Proceedings of the NIPS workshop on cost-sensitive learning. Vancouver, Canada, 2008, pp. 1–10.
-  S. Bianco, L. Celona, P. Napoletano, and R. Schettini, “On the use of deep learning for blind image quality assessment,” Signal, Image and Video Processing, vol. 12, no. 2, pp. 355–362, 2018.
-  S. Bosse, D. Maniry, T. Wiegand, and W. Samek, “A deep neural network for image quality assessment,” in ICIP. IEEE, 2016, pp. 3773–3777.
-  F. Gao, D. Tao, X. Gao, and X. Li, “Learning to rank for blind image quality assessment,” Neural Networks and Learning Systems, IEEE Transactions on, vol. 26, no. 10, pp. 2275–2290, 2015.
-  L. Xu, J. Li, W. Lin, Y. Zhang, L. Ma, Y. Fang, and Y. Yan, “Multi-task rank learning for image quality assessment,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 27, no. 9, pp. 1833–1843, 2017.
-  K. Ma, W. Liu, T. Liu, Z. Wang, and D. Tao, “dipiq: Blind image quality assessment by learning-to-rank discriminable image pairs,” IEEE Transactions on Image Processing, vol. 26, no. 8, pp. 3951–3964, 2017.
-  V. A. Sindagi and V. M. Patel, “A survey of recent advances in cnn-based single image crowd counting and density estimation,” Pattern Recognition Letters, 2017.
-  M. Fu, P. Xu, X. Li, Q. Liu, M. Ye, and C. Zhu, “Fast crowd density estimation with convolutional neural networks,” Engineering Applications of Artificial Intelligence, vol. 43, pp. 81–88, 2015.
-  C. Wang, H. Zhang, L. Yang, S. Liu, and X. Cao, “Deep people counting in extremely dense crowds,” in Proceedings of ACM int. conf. on Multimedia. ACM, 2015.
-  A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in NIPS, 2012, pp. 1097–1105.
-  V. A. Sindagi and V. M. Patel, “Generating high-quality crowd density maps using contextual pyramid cnns,” in ICCV, 2017.
-  C. Zhang, H. Li, X. Wang, and X. Yang, “Cross-scene crowd counting via deep convolutional neural networks,” in CVPR, 2015.
-  D. Kang, D. Dhar, and A. B. Chan, “Crowd counting by adapting convolutional neural networks with side information,” arXiv preprint arXiv:1611.06748, 2016.
-  Y. Li, X. Zhang, and D. Chen, “Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 1091–1100.
-  J. Liu, C. Gao, D. Meng, and A. G. Hauptmann, “Decidenet: Counting varying density crowds through attention guided detection and density estimation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 5197–5206.
-  Z. Shen, Y. Xu, B. Ni, M. Wang, J. Hu, and X. Yang, “Crowd counting via adversarial cross-scale consistency pursuit,” in CVPR, 2018, pp. 5245–5254.
-  Z. Shi, L. Zhang, Y. Liu, X. Cao, Y. Ye, M.-M. Cheng, and G. Zheng, “Crowd counting with deep negative correlation learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 5382–5390.
-  M. Marsde, K. McGuinness, S. Little, C. E. Keogh, and N. E. O’Connor, “People, penguins and petri dishes: adapting object counting models to new visual domains and object types without forgetting,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.
-  V. Ranjan, H. Le, and M. Hoai, “Iterative crowd counting,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018.
-  X. Cao, Z. Wang, Y. Zhao, and F. Su, “Scale aggregation network for accurate and efficient crowd counting,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 734–750.
-  H. Idrees, M. Tayyab, K. Athrey, D. Zhang, S. Al-Maadeed, N. Rajpoot, and M. Shah, “Composition loss for counting, density map estimation and localization in dense crowds,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018.
-  I. H. Laradji, N. Rostamzadeh, P. O. Pinheiro, D. Vazquez, and M. Schmidt, “Where are the blobs: Counting by localization with point supervision,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018.
-  S. Chopra, R. Hadsell, and Y. LeCun, “Learning a similarity metric discriminatively, with application to face verification,” in CVPR, 2005.
-  E. Simo-Serra, E. Trulls, L. Ferraz, I. Kokkinos, P. Fua, and F. Moreno-Noguer, “Discriminative learning of deep convolutional feature point descriptors,” in ICCV, 2015, pp. 118–126.
F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embedding for face recognition and clustering,” inCVPR, 2015, pp. 815–823.
-  X. Wang and A. Gupta, “Unsupervised learning of visual representations using videos,” in ICCV, 2015, pp. 2794–2802.
-  Z. Wang, A. C. Bovik, and L. Lu, “Why is image quality assessment so difficult?” in Acoustics, Speech, and Signal Processing (ICASSP), 2002 IEEE International Conference on, vol. 4. IEEE, 2002, pp. IV–3313.
-  H. R. Sheikh, M. F. Sabir, and A. C. Bovik, “A statistical evaluation of recent full reference image quality assessment algorithms,” Image Processing, IEEE Transactions on, vol. 15, no. 11, pp. 3440–3451, 2006.
-  N. Ponomarenko, O. Ieremeiev, V. Lukin, K. Egiazarian, L. Jin, J. Astola, B. Vozel, K. Chehdi, M. Carli, F. Battisti et al., “Color image database tid2013: Peculiarities and preliminary results,” in Visual Information Processing (EUVIP), 2013 4th European Workshop on. IEEE, 2013, pp. 106–111.
-  H. R. Sheikh, Z. Wang, L. Cormack, and A. C. Bovik, “Live image quality assessment database,” http://live.ece.utexas.edu/research/quality.
-  K. Ma, Q. Wu, Z. Wang, Z. Duanmu, H. Yong, H. Li, and L. Zhang, “Group MAD competition – a new methodology to compare objective image quality models,” in CVPR, 2016.
-  H. Idrees, I. Saleemi, C. Seibert, and M. Shah, “Multi-source multi-scale counting in extremely dense crowd images,” in CVPR, 2013.
-  K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” ICLR, 2015.
-  L. Boominathan, S. S. Kruthiventi, and R. V. Babu, “Crowdnet: a deep convolutional network for dense crowd counting,” in Proc. ACM on Multimedia Conference, 2016.
-  Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for fast feature embedding,” arXiv preprint arXiv:1408.5093, 2014.
-  M. A. Saad, A. C. Bovik, and C. Charrier, “Blind image quality assessment: A natural scene statistics approach in the dct domain,” Image Processing, IEEE Transactions on, vol. 21, no. 8, pp. 3339–3352, 2012.
-  A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain,” Image Processing, IEEE Transactions on, vol. 21, no. 12, pp. 4695–4708, 2012.
-  P. Ye, J. Kumar, L. Kang, and D. Doermann, “Unsupervised feature learning framework for no-reference image quality assessment,” in CVPR. IEEE, 2012, pp. 1098–1105.
-  J. Xu, P. Ye, Q. Li, H. Du, Y. Liu, and D. Doermann, “Blind image quality assessment based on high order statistics aggregation,” IEEE Transactions on Image Processing, vol. 25, no. 9, pp. 4444–4457, 2016.
-  Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004.
-  L. Zhang, L. Zhang, X. Mou, and D. Zhang, “Fsim: a feature similarity index for image quality assessment,” IEEE transactions on Image Processing, vol. 20, no. 8, pp. 2378–2386, 2011.
-  A. K. Moorthy and A. C. Bovik, “Blind image quality assessment: From natural scene statistics to perceptual quality,” IEEE Transactions on Image Processing, vol. 20, no. 12, pp. 3350–3364, 2011.
-  P. Zhang, W. Zhou, L. Wu, and H. Li, “Som: Semantic obviousness metric for image quality assessment,” in CVPR, 2015, pp. 2394–2402.
-  E. Walach and L. Wolf, “Learning to count with cnn boosting,” in ECCV. Springer, 2016.
-  A. B. Chan, Z.-S. J. Liang, and N. Vasconcelos, “Privacy preserving crowd monitoring: Counting people without people models or tracking,” in CVPR, 2008.