Human activity recognition from video is a very active research field, with a long list of potential application domains, ranging from autonomous driving to security surveillance [Aggarwal and Ryoo(2011), Poppe(2010)]. However, the vast majority of published approaches are developed under the assumption that all categories are known a priori [Carreira and Zisserman(2017), Simonyan and Zisserman(2014), Wang et al.(2016)Wang, Xiong, Wang, Qiao, Lin, Tang, and Van Gool, Hara et al.(2018)Hara, Kataoka, and Satoh, Varol et al.(2017)Varol, Laptev, and Schmid, Ji et al.(2013)Ji, Xu, Yang, and Yu]. This closed set
constraint represents a significant bottleneck in the real world, where the system will probably encounter samples from various categories including those never seen during development. The set of possible actions is dynamic by its nature, possibly changing over time. Hence, collecting and maintaining large scale application-specific datasets of video data is especially costly and impractical. This raises a crucial need for the developed models to be able to identify cases where they are faced with samples out of their knowledge domain. In this work, we explore the field of activity recognition underopen set conditions [Scheirer et al.(2014)Scheirer, Jain, and Boult, Scheirer et al.(2013)Scheirer, de Rezende Rocha, Sapkota, and Boult, Busto and Gall(2017)], a setting which has been little-explored before especially in the action recognition domain [Moerland et al.(2016)Moerland, Chandarr, Rudinac, and Jonker].
In an open world application scenario, an action recognition model should be able to handle three different tasks: 1) the standard classification of previously seen categories; 2) knowledge transfer for generalization to new unseen classes (e.g. through zero-shot learning); 3) and knowing how to automatically discriminate between those two cases. The third component of an open set model lies in its ability to identify samples from unseen classes (novelty detection). This is closely linked to the classifier’s confidence in its own predictions, i.e. how can we build models, that know, what they do not know? A straight-forward way is to employ the Softmax
output of a neural network (NN) model as the basis for a rejection threshold[Richter and Roy(2017), Ramos et al.(2017)Ramos, Gehrig, Pinggera, Franke, and Rother]. Traditionally, action recognition algorithms focus on maximizing the top-1 performance on a static set of actions. Such optimization leads to Softmax scores of the winning class being strongly biased towards very high values [Nguyen et al.(2015)Nguyen, Yosinski, and Clune, Szegedy et al.(2013)Szegedy, Zaremba, Sutskever, Bruna, Erhan, Goodfellow, and Fergus, Gal and Ghahramani(2016), Kendall et al.(2017)Kendall, Gal, and Cipolla]
. While giving excellent results in closed set classification, such overly self-confident models become a burden under open set conditions. A better way to asses NN’s confidence, is to rather predict the probability distribution with Bayesian neural networks (BNN). Recently, Galet al [Gal and Ghahramani(2016)] introduced a way of efficiently approximating BNN modeled as a Gaussian Process [Rasmussen(2004)] and using dropout-based Monte-Carlo sampling (MC-Dropout) [Gal and Ghahramani(2016)]. We leverage the findings of [Gal and Ghahramani(2016)] and exploit the predictive uncertainty in order to identify activities of previously unseen classes.
This work aims at bringing conventional activity recognition to a setting where new categories might occur at any time and has the following main contributions: 1) We present a new model for novelty detection for action recognition based on the predictive uncertainty of the classifiers. Our main idea is to estimate the novelty of a new sample based on the uncertainty of a selected group of output classifiers in a voting-like manner. The choice of the voting classifiers depends on how confident they are in relation to the currently predicted class. 2) We adapt zero-shot action recognition models, which are conventionally applied solely on samples of the unseen classes, to the generalized case (i.e. open set scenario) where a test sample may originate from either known or novel categories. We present a generic framework for generalized zero-shot action recognition, where our novelty detection model serves as a filter to distinguish between seen and novel categories, passing the sample either to a standard classifier or a zero-shot model accordingly. 3) We extend the custom evaluation setup for action recognition to the open-set scenario and formalize the evaluation protocol for the tasks of novelty detection and zero-shot action recognition in the generalized case on two well-established datasets, UCF-101 [Soomro et al.(2012)Soomro, Zamir, and Shah] and HMDB-51 [Kuehne et al.(2013)Kuehne, Jhuang, Stiefelhagen, and Serre]. The evaluation shows, that our model consistently outperforms conventional NNs and other baseline methods in identifying novel activities and was highly successful when applied to generalized zero-shot learning.
2 Related Work
Various machine learning methods have been used for quantifying thenormality of a data sample. An overview of the existing approaches is provided by [Pimentel et al.(2014)Pimentel, Clifton, Clifton, and Tarassenko, Chandola et al.(2009)Chandola, Banerjee, and Kumar]. A lot of today’s novelty detection research is handled from the probabilistic point of view [Liu et al.(2017)Liu, Lian, Wang, and Xiao, Socher et al.(2013)Socher, Ganjoo, Manning, and Ng, Pimentel et al.(2014)Pimentel, Clifton, Clifton, and Tarassenko, Mohammadi-Ghazi et al.(2018)Mohammadi-Ghazi, Marzouk, and Büyüköztürk]Pimentel et al.(2014)Pimentel, Clifton, Clifton, and Tarassenko]. The One-class SVM introduced by Schölkopf et al[Schölkopf et al.(2001)Schölkopf, Platt, Shawe-Taylor, Smola, and Williamson]
is another widely used unsupervised method for novelty detection, mapping the training data into the feature space and maximizing the margin of separation from the origin. Anomaly detection with NNs has been addressed several times using encoder-decoder-like architectures and the reconstruction error[Williams et al.(2002)Williams, Baxter, He, Hawkins, and Gu]
. A common way for anomaly detection is to threshold the output of the neuron with the highest value[Markou and Singh(2006), Hendrycks and Gimpel(2017), Richter and Roy(2017)]. Recently, Hendrycks et al[Hendrycks and Gimpel(2017)]
The research of novelty detection in videos has been very limited. A related topic of anomaly detection has been studied for very specific applications, such as surveillance [Pimentel et al.(2014)Pimentel, Clifton, Clifton, and Tarassenko, Markou and Singh(2006)] or personal robotics[Moerland et al.(2016)Moerland, Chandarr, Rudinac, and Jonker]. Surveillance however often has anomalies, such as Robbery or Vandalism, present in the training set in some form [Mohammadi et al.(2016)Mohammadi, Perina, Kiani, and Murino, Sultani et al.(2018)Sultani, Chen, and Shah] which violates our open-set assumption. The work most similar to ours is the one of Moerland et al[Moerland et al.(2016)Moerland, Chandarr, Rudinac, and Jonker]
where Hidden-Markov-Model is used to detect unseen actions from skeleton features. However,[Moerland et al.(2016)Moerland, Chandarr, Rudinac, and Jonker] considers only a simplified evaluation setting using only a single unseen action category in testing. In contrast to [Moerland et al.(2016)Moerland, Chandarr, Rudinac, and Jonker] our model is based on a deep neural architecture for detecting novel actions which makes it applicable to a wide range of modern action recognition models. Furthermore, we consider a challenging evaluation setting on well-established datasets where novel classes are as diverse as those seen before. Additionally, we go beyond novelty detection and evaluate how well our model generalizes to classifying novel classes through zero-shot learning. Our model leverages approximation of BNN using MC-Dropout as proposed by Gal et al[Gal and Ghahramani(2016)], which has been successfully applied in semantic segmentation [Kendall et al.(2017)Kendall, Gal, and Cipolla]
and active learning[Gal et al.(2017)Gal, Islam, and Ghahramani]. We extend the BNN approximation to the context of open set action recognition where we incorporate the uncertainty of the output neurons in a voting scheme for novelty detection.
Zero-Shot Action Recognition
Research on human activity recognition under open set conditions has been sparse so far. A related field of Zero-Shot Learning (ZSL) attempts to classify new actions without any training data by linking visual features and the high-level semantic descriptions of a class, e.g
through action labels. The description is often represented with word vectors by a skip-gram model (e.gword2vec [Mikolov et al.(2013)Mikolov, Sutskever, Chen, Corrado, and Dean]) previously trained on a large-scale text corpus. ZSL for action recognition gained popularity over the past few years and has also been improving slowly but steadily [Xu et al.(2017)Xu, Hospedales, and Gong, Wang and Chen(2017), Xu et al.(2015)Xu, Hospedales, and Gong, Qin et al.(2017)Qin, Liu, Shao, Shen, Ni, Chen, and Wang, Zhu et al.(2018)Zhu, Long, Guan, Newsam, and Shao]. In all of these works, the categories used for training and testing are disjoint and the method is evaluated on unfamiliar actions only. This is not a realistic scenario, since it requires the knowledge of whether the activity belongs to a known or novel category a priori. Generalized zero-shot learning (GZSL) has been recently studied for image recognition and a drastic performance drop of classical ZSL approaches such as ConSE [Norouzi et al.(2013)Norouzi, Mikolov, Bengio, Singer, Shlens, Frome, Corrado, and Dean] and Devise [Frome et al.(2013)Frome, Corrado, Shlens, Bengio, Dean, Mikolov, et al.], has been reported [Xian et al.(2017)Xian, Schiele, and Akata]. As the main application of our novelty detection approach, we implement a framework for ZSL in the generalized case and integrate our novelty detection method to distinguish between known and unknown actions.
3 Novelty Detection via Informed Voting
We present a new approach for novelty detection in action recognition. That is, given a new video sample , our goal is to find out whether is a sample of a previously known category or if it belongs to a novel action category not seen before during training.
Let be the set of all known categories in our dataset. Then is the classifier probability of action category given sample . Conceptually, our novelty detection model is composed of two main components: 1) the leader and 2) the council. The leader refers to the classifier with the highest confidence score in predicting the class of a certain sample . For example, in classification neural networks it is common to select the leader based on the highest softmax prediction score. The leader votes for sample being of its own category and assuming that the class of is one of the known categories, i.e. . The council, on the other hand, is a subset of classifiers that will help us validating the decision of a specific leader. In other words, the council members of a leader representing the selected class are a subset of the classifiers representing the rest of the classes, i.e. . These members are elected for each leader individually, i.e.
each category classifier in our model has its own council. A council member is selected based on its certainty variance in relation to a leader. Whenever a leader decides on the category of a sample, its council will convene and vote on the leader decision. Then, the council members will jointly decide whether the leader made the correct decision or it was mistaken because the sample is actually from a novel category.
Next, we explain in details how we measure the uncertainty of a classifier (Section 3.1); choosing a leader and its council members (Section 3.2); and, finally, the novelty voting procedure given new sample (Section 3.3).
3.1 Measuring Classifier Uncertainty
In this section, we tackle the problem of quantifying the uncertainty of a classifier given a new sample. The estimated uncertainty is leveraged later by our model to select the council members as we will see in Section 3.2.
In the context of deep learning, it is common to consider the single point estimates for each category, represented by the output of the softmax layer, as a confidence measure[Markou and Singh(2006), Hendrycks and Gimpel(2017), Augusteijn and Folkert(2002), Richter and Roy(2017)]. However, this practice has been highlighted in literature to be inaccurate since a model can be highly uncertain even when producing high prediction scores [Nguyen et al.(2015)Nguyen, Yosinski, and Clune, Szegedy et al.(2013)Szegedy, Zaremba, Sutskever, Bruna, Erhan, Goodfellow, and Fergus]. Bayesian neural networks (BNNs) offer us an alternative to the point estimate models and are known to provide a well calibrated estimation of the network uncertainty in its output. Given the network parameters and a training set , the predictive probability of the BNN is obtained by integrating over the parameter space. The prediction
is therefore the mean over all possible parameter combinations weighted by their posterior probability:
However, BNNs are known to have a difficult inference scheme and high computation cost [Gal and Ghahramani(2016)]. Therefore, we leverage the robust model proposed by [Gal and Ghahramani(2016)] to approximate the predictive mean and uncertainty of the BNN posterior distribution with network parameters modeled as a Gaussian Process (GP). This method is based on dropout regularization [Srivastava et al.(2014)Srivastava, Hinton, Krizhevsky, Sutskever, and Salakhutdinov]
, a widely used technique which has proven to be very effective against overfitting. That is, it leverages the dropout at each layer in the network to draw the weights from a Bernoulli distribution with probability. At test time, the dropout is iteratively applied for forward passes for each individual sample. Then, the statistics of the neural network output represents a Monte-Carlo (MC) approximation of the neuron’s posterior distribution This approach is referred to as MC-Dropout [Gal and Ghahramani(2016)].
be a representation generated by a convolutional neural network (CNN) for an input sample. We add a feedforward network on top of the CNN with two fully-connected layers with weight matrices and . Instead of using a deterministic Softmax estimate in a single forward pass as it is common with CNNs, we now compute the mean over stochastic iterations as our prediction score:
is the bias vector of the firs layer. Additionally,and are diagonal matrices where the diagonal elements contain binary values, such that they are set to with probability and otherwise to .
We further empirically compute the model’s predictive uncertainty as the distribution variance:
Fig. 1 shows how predictive mean and uncertainty are distributed for samples of known and novel classes. The plot depicts clearly different patterns for the resulting probability distributions in these two cases which illustrates the potential of Bayesian uncertainty for novelty detection.
3.2 Selecting the Leader and its Council
Now that we can estimate the confidence and uncertainty of each category classifier in our model, we describe in this section how to choose the leader and select it council members.
Rather than selecting the leader using a point estimate based on the softmax scores of the output layer, we leverage here the more stable dropout-based estimation of the prediction mean. Hence, the leader is selected as the classifier with highest expected prediction score over sampling iterations:
where is estimated according to Eq. 2.
The leader by itself can sometimes produce highly confident predictions for samples of unseen categories [Szegedy et al.(2013)Szegedy, Zaremba, Sutskever, Bruna, Erhan, Goodfellow, and Fergus]. Hence, we can not rely solely on the leader confidence to estimate whether a sample is of a novel category or not. Here, the rest of the classifiers can help in checking the validity of the leader’s decision. We notice that these classifier exhibit unique patterns in regard to a certain leader. They can be grouped into two main groups: the first shows high uncertainty when the leader is correctly classifying a sample; while the second shows a very low uncertainty and are in agreement with the leader.
Guided by this observation, we select the members of the Council for a certain leader based on their uncertainty variance in regards to samples of the leader’s category, i.e. . In other words, those classifiers that exhibit very low uncertainty when the leader is classifying samples of its own category are elected to join its council. During the training phase, we can select the council members for each classifier in our model. Here, we randomly split the initial set into a training set which is used for model optimization and parameter estimation, and a holdout set which is used for choosing the council member for all the classifiers iteratively. Specifically, we use a 9/1 split for the training and the holdout splits. We first estimate the parameters of our deep model using . Then, we evaluate our model over all samples from . For each category classifier in our model, we construct a set of true positive samples . For each sample , we estimate the uncertainty of the rest of the classifiers using the MC-Dropout approach. Then, the variance of these classifiers’ uncertainty is estimated as:
where and is the expectation of the uncertainty of the classifier over samples . Finally, classifiers with a variance lower than a fixed credibility threshold are then elected as members of council.
Fig. 2 shows three leaders and their elected councils according to our approach. We see, for example, that eight classifiers did not pass the credibility threshold for the leader drink and were excluded from its council. The variance of the uncertainty is especially high for sit and eat in this case. This is expected since those actions often occur in a similar context.
3.3 Voting for Novelty
Given the trained deep model and the sets of all council members from the previous step, we can now generate a novelty score for a new sample as follows. First, we calculate the prediction mean and uncertainty of all the action classifiers using stochastic forward passes and MC-Dropout. Then, the classifier with the maximum predicted mean is chosen as the leader. Finally, the council members of the chosen leader vote for the novelty of sample based on their estimated uncertainty (see Algorithm 3.3).
Examples of such voting outcome for three different leaders are illustrated in Fig. 3. In case of category cartwheel, we can see that when the leader is voting indeed for the correct category, all council members show low uncertainty values therefore resulting in a low novelty score, as uninformed classifiers (marked in red) are excluded. However, we observe very different measurements for an example from an unseen category clap which is also predicted as cartwheel. Here, multiple classifiers which are in the council (marked in blue) show unexpected high uncertainty values (e.g. eat, laugh), therefore discrediting the leader decision and voting for a high novelty score.
We refer to our previous model as the Informed Democracy model since voting is restricted to the council members which are chosen in an informed manner to check the decision of the leader. In addition to the previous model, we consider two other variants of our model:
The Uninformed Democracy model: Here, there is no council and all classifiers have the right to vote for any leader. Hence, step in Algorithm 3.3 is replaced with .
The Dictator model: unlike the previous model, this one leverages only the leader’s uncertainty in its own decision to predict the novelty of the sample, i.e. .
Open set and zero-shot learning
Once our model generated the novelty score , we can decide whether is a sample from a novel category or not using a sensitivity threshold
. This threshold can be estimated from a validation set using the equal error rate of the receiver operating characteristic curve (ROC). Then, ifthe Council votes in favor of the Leader and its category is taken as our final classification result. Otherwise, an unknown activity class has been identified. In this case, the input could be passed further to a module in charge of handling unfamiliar data, such as a zero-shot learning model or a user to give the sample a new label in the context of active learning.
Since there is no established evaluation procedure available for action recognition in open-set conditions, we adapt existing evaluation protocols for two well-established datasets, HMDB-51 [Kuehne et al.(2013)Kuehne, Jhuang, Stiefelhagen, and Serre] and UCF-101 [Soomro et al.(2012)Soomro, Zamir, and Shah], for our task222Dataset splits used for novelty detection and generalized zero-shot action recognition are provided at https://cvhci.anthropomatik.kit.edu/aroitberg/novelty_detection_action_recognition. We evenly split each dataset into seen/unseen categories (26/25 for HMDB-51 and 51/50 for UCF-101). Samples of unseen classes will not be available during training, while samples of the remaining set of seen classes is further split into training (70%) and testing (30%) sets, thereby adapting the evaluation framework of [Wang and Chen(2017)] for the generalized
ZS learning scenario. For each dataset, we randomly generate 10 splits and report the average and standard deviation of the recognition accuracy. Using a separate validation split, we optimize the credibility thresholdand compute the threshold for rejection for each category as the Equal Error Rate of the ROC.
. The last average pooling is connected to two fully connected layers: a hidden layer of size 256 and the final softmax-classifier layer. These are optimized using SGD with momentum of 0.9, learning rate of 0.005 and dropout probability of 0.7 for 100 epochs. We sample the output scores forstochastic forward passes applied on the two layers preceding the classifier, while the credibility threshold is set to 0.001.
We compare our model to three popular methods for novelty and outlier detection: 1) a One Class SVM[Schölkopf et al.(2000)Schölkopf, Williamson, Smola, Shawe-Taylor, and Platt, Schölkopf et al.(2001)Schölkopf, Platt, Shawe-Taylor, Smola, and Williamson] with RBF kernel (upper bound on the fraction of training errors set to 0.1); 2) a GMM [Zorriassatine et al.(2005)Zorriassatine, Al-Habaibeh, Parkin, Jackson, and Coy, Pimentel et al.(2014)Pimentel, Clifton, Clifton, and Tarassenko] with 8 components; 3) and Softmax probabilities [Richter and Roy(2017), Hendrycks and Gimpel(2017)] as the value for thresholding. Both SVM and GMM were trained on normalized features obtained from last average pooling layer of I3D pre-trained on the Kinetics dataset [Carreira and Zisserman(2017)].
We evaluate the novelty detection accuracy in terms of a binary classification problem, using the area under curve (AUC) values of the receiver operating characteristic (ROC) and the precision-recall (PR) curves.
We show the robustness of our approach in comparison to the baseline methods in Table 1. All variants of our model clearly outperform the conventional approaches and achieve an ROC-AUC gain of over 7% on both datasets. Along our model variants, Informed Democracy has proven to be the most effective strategy for novelty score voting, outperforming the Dictator by 5.5% and 1.4%, while Uninformed Democracy achieved second-best results. We believe that smaller differences in performance gain on the UCF-101 data are due to the much higher supervised classification accuracy on this dataset. Since the categories of UCF-101 are easier to distinguish visually and the confusion is low, there is more agreement between the neurons in terms of their confidence.
Generalized Zero-Shot Learning (GZSL)
Next, we evaluate our approach in the context of GZSL, where our novelty detection model serves as a filter to distinguish whether the observed example should be classified with the I3D model in the standard classification setup, or mapped to one of the unknown classes via a ZSL model. We compare two prominent ZSL methods: ConSE [Norouzi et al.(2013)Norouzi, Mikolov, Bengio, Singer, Shlens, Frome, Corrado, and Dean] and DeViSE [Frome et al.(2013)Frome, Corrado, Shlens, Bengio, Dean, Mikolov, et al.]. The ConSE model starts by predicting probabilities of the seen classes, and then takes the convex combination of word embeddings of the top K most possible seen classes and select its nearest neighbor from the novel classes in the word2vec space. For DeViSE, we train a separate model to regress word2vec representations from the visual features. We use the publicly available word2vec model that is trained on Google News articles [Mikolov et al.(2013)Mikolov, Sutskever, Chen, Corrado, and Dean].
For consistency, we first report the results for the standard ZS case (i.e. UU) and further extend to the generalized case (i.e. UU+S and U+SU+S) as shown in Table 2. In the more realistic GZSL setup, our model is not restricted to any group of target labels and is evaluated on actions of seen and unseen category using the harmonic mean of accuracies for seen and unseen classes as proposed by [Xian et al.(2017)Xian, Schiele, and Akata]. Table 2 shows a clear advantage of employing novelty detection as part of a GZSL framework. While failure of the original ConSE and DeViSE models might be surprising at first glance, such performance drops have been discussed in previous work on ZSL for image recognition [Xian et al.(2017)Xian, Schiele, and Akata] and is due to the fact that both models are biased towards labels that were used during training. Our Informed Democracy model yields the best recognition rates in every setting and can therefore be indeed successfully applied for multi-label action classification in case of new activities.
|Novelty Detection Model||HMDB-51||UCF-101|
|ROC AUC %||PR AUC %||ROC AUC %||PR AUC %|
|One-class SVM||54.09 (3.0)||77.86 (4.0)||53.55 (2.0)||78.57 (2.4)|
|Gaussian Mixture Model||56.83 (4.2)||78.40 (3.6)||59.21 (4.2)||79.50 (2.2)|
|Conventional NN Confidence||67.58 (3.3)||84.21 (3.0)||84.28 (1.9)||93.92 (0.7)|
|Our Proposed Model based on Bayesian Uncertainty|
|Dictator||71.78 (1.8)||86.81 (2.5)||91.43 (2.3)||96.72 (1.0)|
|Uninformed Democracy||73.81 (1.7)||87.83 (2.3)||92.13 (1.8)||97.15 (0.7)|
|Informed Democracy||75.33 (2.7)||88.66 (2.3)||92.94 (1.7)||97.52 (0.6)|
|Standard ConSe Model||21.03 (2.07)||0 (0)||0 (0)||17.85 (1.95)||0.07 (0.10)||0.13 (0.20)|
|Standard Devise Model||17.27 (2.01)||0.26 (0.37)||0.52 (0.73)||14.48 (1.13)||0.81 (0.36)||1.61 (0.71)|
|ConSe + Novelty Detection|
|One-class SVM||21.03 (2.07)||10.99 (1.83)||17.40 (2.41)||17.85 (1.95)||10.37 (1.59)||16.55 (1.91)|
|Gaussian Mixture Model||21.03 (2.07)||13.30 (2.58)||19.91 (3.32)||17.85 (1.95)||9.31 (1.30)||15.98 (1.99)|
|Conventional NN Confidence||21.03 (2.07)||10.96 (0.87)||18.56 (1.22)||17.85 (1.95)||12.19 (1.72)||20.91 (2.59)|
|Informed Democracy (ours)||21.03 (2.07)||13.67 (1.31)||22.27 (1.79)||17.85 (1.95)||13.62 (1.94)||23.42 (2.97)|
|Devise + Novelty Detection|
|One-class SVM||17.27 (2.01)||8.92 (1.89)||14.67 (2.74)||14.48 (1.13)||8.65 (1.59)||14.25 (2.00)|
|Gaussian Mixture Model||17.27 (2.01)||10.61 (2.22)||16.72 (3.1)||14.48 (1.13)||7.26 (0.84)||12.88 (1.40)|
|Conventional NN Confidence||17.27 (2.01)||8.68 (1)||15.17 (1.56)||14.48 (1.13)||10.08 (1.59)||17.69 (2.33)|
|Informed Democracy (ours)||17.27 (2.01)||10.73 (1.47)||18.18 (2.21)||14.48 (1.13)||11.03 (1.42)||19.48 (2.21)|
We introduce a new approach for novelty detection in action recognition. Our model leverages the estimated uncertainty of the category classifiers to detect samples from novel categories not encountered during training. This is achieved by selecting a council of classifiers for each leader (i.e. the most confident classifier). The council will validate the decision made by the leader through voting. Hence, either confirming the classification decision for a sample of a known category or revoking the leader decision and deeming the sample to be novel. We show in a thorough evaluation on two challenging benchmark, that our model outperforms the state-of-the-art in novelty detection. Furthermore, we demonstrate that our model can be easily integrated in a generalized zero-shot learning framework. Combining our model with off-the-shelf zero-shot approaches leads to significant improvements in classification accuracy.
This work has been partially funded by the German Federal Ministry of Education and Research (BMBF) within the PAKoS project.
- [Aggarwal and Ryoo(2011)] Jake K Aggarwal and Michael S Ryoo. Human activity analysis: A review. ACM Computing Surveys (CSUR), 43(3):16, 2011.
- [Augusteijn and Folkert(2002)] MF Augusteijn and BA Folkert. Neural network classification and novelty detection. International Journal of Remote Sensing, 23(14):2891–2902, 2002.
- [Busto and Gall(2017)] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In The IEEE International Conference on Computer Vision (ICCV), volume 2, 2017.
[Carreira and Zisserman(2017)]
Joao Carreira and Andrew Zisserman.
Quo vadis, action recognition? a new model and the kinetics dataset.
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4724–4733. IEEE, 2017.
- [Chandola et al.(2009)Chandola, Banerjee, and Kumar] Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. ACM computing surveys (CSUR), 41(3):15, 2009.
- [Frome et al.(2013)Frome, Corrado, Shlens, Bengio, Dean, Mikolov, et al.] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Tomas Mikolov, et al. Devise: A deep visual-semantic embedding model. In Advances in neural information processing systems, pages 2121–2129, 2013.
- [Gal and Ghahramani(2016)] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050–1059, 2016.
- [Gal et al.(2017)Gal, Islam, and Ghahramani] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian Active Learning with Image Data. In Proceedings of the 34th International Conference on Machine Learning (ICML-17), 2017.
[Hara et al.(2018)Hara, Kataoka, and Satoh]
Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh.
Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet.In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, pages 18–22, 2018.
- [Hendrycks and Gimpel(2017)] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In Proceedings of International Conference on Learning Representations, 2017.
- [Ji et al.(2013)Ji, Xu, Yang, and Yu] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221–231, 2013.
- [Kendall et al.(2017)Kendall, Gal, and Cipolla] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics. In arXiv:1705.07115, May 2017.
- [Kuehne et al.(2013)Kuehne, Jhuang, Stiefelhagen, and Serre] Hilde Kuehne, Hueihan Jhuang, Rainer Stiefelhagen, and Thomas Serre. Hmdb51: A large video database for human motion recognition. In High Performance Computing in Science and Engineering ‘12, pages 571–582. Springer, 2013.
- [Liu et al.(2017)Liu, Lian, Wang, and Xiao] Juncheng Liu, Zhouhui Lian, Yi Wang, and Jianguo Xiao. Incremental kernel null space discriminant analysis for novelty detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 792–800, 2017.
- [Markou and Singh(2006)] Markos Markou and Sameer Singh. A neural network-based novelty detector for image sequence analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(10):1664–1677, 2006.
- [Mikolov et al.(2013)Mikolov, Sutskever, Chen, Corrado, and Dean] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119, 2013.
- [Moerland et al.(2016)Moerland, Chandarr, Rudinac, and Jonker] Thomas Moerland, Aswin Chandarr, Maja Rudinac, and Pieter P Jonker. Knowing what you don’t know-novelty detection for action recognition in personal robots. In VISIGRAPP (4: VISAPP), pages 317–327, 2016.
- [Mohammadi et al.(2016)Mohammadi, Perina, Kiani, and Murino] Sadegh Mohammadi, Alessandro Perina, Hamed Kiani, and Vittorio Murino. Angry crowds: detecting violent events in videos. In European Conference on Computer Vision, pages 3–18. Springer, 2016.
- [Mohammadi-Ghazi et al.(2018)Mohammadi-Ghazi, Marzouk, and Büyüköztürk] Reza Mohammadi-Ghazi, Youssef M Marzouk, and Oral Büyüköztürk. Conditional classifiers and boosted conditional gaussian mixture model for novelty detection. Pattern Recognition, 2018.
- [Nguyen et al.(2015)Nguyen, Yosinski, and Clune] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 427–436, 2015.
- [Norouzi et al.(2013)Norouzi, Mikolov, Bengio, Singer, Shlens, Frome, Corrado, and Dean] Mohammad Norouzi, Tomas Mikolov, Samy Bengio, Yoram Singer, Jonathon Shlens, Andrea Frome, Greg S Corrado, and Jeffrey Dean. Zero-shot learning by convex combination of semantic embeddings. arXiv preprint arXiv:1312.5650, 2013.
- [Pimentel et al.(2014)Pimentel, Clifton, Clifton, and Tarassenko] Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. A review of novelty detection. Signal Processing, 99:215–249, 2014.
- [Poppe(2010)] Ronald Poppe. A survey on vision-based human action recognition. Image and vision computing, 28(6):976–990, 2010.
- [Qin et al.(2017)Qin, Liu, Shao, Shen, Ni, Chen, and Wang] Jie Qin, Li Liu, Ling Shao, Fumin Shen, Bingbing Ni, Jiaxin Chen, and Yunhong Wang. Zero-shot action recognition with error-correcting output codes. In Proc. CVPR, 2017.
- [Ramos et al.(2017)Ramos, Gehrig, Pinggera, Franke, and Rother] Sebastian Ramos, Stefan Gehrig, Peter Pinggera, Uwe Franke, and Carsten Rother. Detecting unexpected obstacles for self-driving cars: Fusing deep learning and geometric modeling. In Intelligent Vehicles Symposium (IV), 2017 IEEE, pages 1025–1032. IEEE, 2017.
- [Rasmussen(2004)] Carl Edward Rasmussen. Gaussian processes in machine learning. In Advanced lectures on machine learning, pages 63–71. Springer, 2004.
- [Richter and Roy(2017)] Charles Richter and Nicholas Roy. Safe visual navigation via deep learning and novelty detection. In Proc. of the Robotics: Science and Systems Conference, 2017.
- [Scheirer et al.(2013)Scheirer, de Rezende Rocha, Sapkota, and Boult] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(7):1757–1772, 2013.
- [Scheirer et al.(2014)Scheirer, Jain, and Boult] Walter J Scheirer, Lalit P Jain, and Terrance E Boult. Probability models for open set recognition. IEEE transactions on pattern analysis and machine intelligence, 36(11):2317–2324, 2014.
- [Schölkopf et al.(2000)Schölkopf, Williamson, Smola, Shawe-Taylor, and Platt] Bernhard Schölkopf, Robert C Williamson, Alex J Smola, John Shawe-Taylor, and John C Platt. Support vector method for novelty detection. In Advances in neural information processing systems, pages 582–588, 2000.
- [Schölkopf et al.(2001)Schölkopf, Platt, Shawe-Taylor, Smola, and Williamson] Bernhard Schölkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. Estimating the support of a high-dimensional distribution. Neural computation, 13(7):1443–1471, 2001.
- [Simonyan and Zisserman(2014)] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in neural information processing systems, pages 568–576, 2014.
- [Socher et al.(2013)Socher, Ganjoo, Manning, and Ng] Richard Socher, Milind Ganjoo, Christopher D Manning, and Andrew Ng. Zero-shot learning through cross-modal transfer. In Advances in neural information processing systems, pages 935–943, 2013.
- [Soomro et al.(2012)Soomro, Zamir, and Shah] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.
- [Srivastava et al.(2014)Srivastava, Hinton, Krizhevsky, Sutskever, and Salakhutdinov] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958, 2014.
- [Sultani et al.(2018)Sultani, Chen, and Shah] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. arXiv preprint arXiv:1801.04264, 2018.
- [Szegedy et al.(2013)Szegedy, Zaremba, Sutskever, Bruna, Erhan, Goodfellow, and Fergus] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013.
- [Varol et al.(2017)Varol, Laptev, and Schmid] Gul Varol, Ivan Laptev, and Cordelia Schmid. Long-term temporal convolutions for action recognition. IEEE transactions on pattern analysis and machine intelligence, 2017.
- [Wang et al.(2016)Wang, Xiong, Wang, Qiao, Lin, Tang, and Van Gool] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European Conference on Computer Vision, pages 20–36. Springer, 2016.
- [Wang and Chen(2017)] Qian Wang and Ke Chen. Zero-shot visual recognition via bidirectional latent embedding. International Journal of Computer Vision, 124(3):356–383, 2017.
- [Williams et al.(2002)Williams, Baxter, He, Hawkins, and Gu] Graham Williams, Rohan Baxter, Hongxing He, Simon Hawkins, and Lifang Gu. A comparative study of rnn for outlier detection in data mining. In Data Mining, 2002. ICDM 2003. Proceedings. 2002 IEEE International Conference on, pages 709–712. IEEE, 2002.
- [Xian et al.(2017)Xian, Schiele, and Akata] Yongqin Xian, Bernt Schiele, and Zeynep Akata. Zero-shot learning-the good, the bad and the ugly. arXiv preprint arXiv:1703.04394, 2017.
- [Xu et al.(2015)Xu, Hospedales, and Gong] Xun Xu, Timothy Hospedales, and Shaogang Gong. Semantic embedding space for zero-shot action recognition. In Image Processing (ICIP), 2015 IEEE International Conference on, pages 63–67. IEEE, 2015.
- [Xu et al.(2017)Xu, Hospedales, and Gong] Xun Xu, Timothy Hospedales, and Shaogang Gong. Transductive zero-shot action recognition by word-vector embedding. International Journal of Computer Vision, pages 1–25, 2017.
- [Zhu et al.(2018)Zhu, Long, Guan, Newsam, and Shao] Yi Zhu, Yang Long, Yu Guan, Shawn Newsam, and Ling Shao. Towards universal representation for unseen action recognition. 2018.
- [Zorriassatine et al.(2005)Zorriassatine, Al-Habaibeh, Parkin, Jackson, and Coy] F Zorriassatine, A Al-Habaibeh, RM Parkin, MR Jackson, and J Coy. Novelty detection for practical pattern recognition in condition monitoring of multivariate processes: a case study. The International Journal of Advanced Manufacturing Technology, 25(9-10):954–963, 2005.