Successful action recognition will help us drive our cars, prevent crime, search our video collections and will eventually enable robots to serve us at home. Such applications require action localization, i.e. identifying when the action happens and who is performing the action. Most of the current methods and benchmarks for action recognition, however, only address action classification de Souza et al. (2016); Simonyan and Zisserman (2014), i.e. assuming temporally segmented action intervals as input.
Identifying the beginning and the end of an action naturally suggests the need of temporal models for video sequences. Sequence models have previously been explored for sound, speech and text understanding. In particular, recurrent neural network models (RNNs) have recently shown success for speech recognitionDahl et al. (2012)
and text generationSutskever et al. (2011) as well as for image and video captioning Donahue et al. (2015); Karpathy and Fei-Fei (2015); Vinyals et al. (2015). RNNs have also been explored for action classification in video Donahue et al. (2015); Ng et al. (2015), but have shown limited improvements for this task so far.
Action classification may not require sophisticated temporal models if classes can be distinguished solely by the presence of action-specific features. On the other hand, if the structure of the video is required as the output, explicit spatio-temporal models of the video can be beneficial. Recent work Liu et al. (2016a); Ma et al. (2016); Singh et al. (2016); Yeung et al. (2016); Yuan et al. (2016) has indeed shown improvements in temporal action localization achieved with recurrent models of video sequences. Here we develop and investigate recurrent models for spatio-temporal action localization.
Our goal is to localize the acting person in the video frame and to identify temporal boundaries of corresponding actions. To this end, we propose a recurrent localization network (RecLNet) with gated recurrent units (GRU) Cho et al. (2014) for modeling actions on the level of person tracks. Our method starts by person detection and tracking similar to Gkioxari and Malik (2015); Peng and Schmid (2016); Saha et al. (2016); Weinzaepfel et al. (2015). Differently to previous work, we train our RecLNet to score actions and detect temporal boundaries within each person track (see Figure 1). This scoring is achieved by two-stream recurrent units exploiting appearance and motion Fast-RCNN Girshick (2015) features pooled from person boxes, while final detections are obtained using our simple and effective temporal localization method composed of filtering and thresholding. We provide a thorough experimental evaluation and analyze the impact of recurrence on spatio-temporal action localization by making the following contributions:
we show spatio-temporal action localization improvement supported by our RecLNet trained on a track-level and compare different standard and recurrent architectures;
we empirically diagnose temporal localization as being a weakness of existing methods which our method is able to correct;
our method is complementary to most recent works Kalogeiton et al. (2017) mainly focusing on increasing spatial boxes precision and we identify the spatial aspect as being our principal room for improvement;
The rest of the paper is organized as follows. Section 2 reviews the related work on action classification, temporal and spatio-temporal localization. Section 3 introduces our RecLNet model, its architecture and our threshold temporal localization technique. Section 5 presents our experimental and qualitative results. Section 6 finally draws conclusions.
2 Related work
Our work is mostly related to methods for human action classification, temporal action localization and spatio-temporal action detection in video.
Action classification. The majority of action recognition methods targets clip-level video classification, i.e. the assignment of video clips to a closed set of action classes. Recent datasets for this task include UCF-101 Soomro et al. (2012), HMDB Kuehne et al. (2011), ActivityNet Heilbron et al. (2015) and Sports-1M Karpathy et al. (2014). Local space-time features such as HOG, HOF and IDT Laptev et al. (2008); Wang and Schmid (2013) have shown initial progress for this task. More recently CNN and RNN-based approaches have been investigated to learn video representations for action recognition. A combination of motion and appearance information learned by two separate CNN networks has been proposed in Simonyan and Zisserman (2014). Alternative methods based on spatio-temporal convolutions have been studied in Ji et al. (2010); Taylor et al. (2010) and more recently using the C3D Tran et al. (2015) and I3D Carreira and Zisserman (2017a) architectures. Such method Carreira and Zisserman (2017a)
takes advantage of transfer learning by training 3D architectures on a large video datasetsKay et al. (2017a). Several works used RNNs for aggregating video information along time Baccouche et al. (2011); Donahue et al. (2015); Ng et al. (2015); Pigou et al. (2017); Singh et al. (2016); Srivastava et al. (2016). Currently best performing action classification methods combine IDT features with CNN-based representations of motion and appearance de Souza et al. (2016); Feichtenhofer et al. (2016a, b); Varol et al. (2017) or pure 3D CNN architecture Carreira and Zisserman (2017a). RNNs have shown promise for the task of gesture recognition in Pigou et al. (2017)
but did not show significant improvements for the more general task of action classification. In this work, we compute appearance and motion CNN features extracted on the level of person tracks and use them as input to our RecLNet for action localization.
Temporal action localization aims both to classify and identify temporal extents of actions in longer video clips. Recent datasets for this task include THUMOSIdrees et al. (2016), Activity-Net Heilbron et al. (2015) and MPII Cooking Rohrbach et al. (2016). Methods for joint action classification and temporal segmentation have explored dynamic programming Hoai et al. (2011) and temporal grammars Pirsiavash and Ramanan (2014). More recently several RNN-based methods have shown gains for action localization in Bagautdinov et al. (2017); Ma et al. (2016); Singh et al. (2016); Yeung et al. (2016); Yuan et al. (2016). For example, Ma et al. (2016); Singh et al. (2016); Yuan et al. (2016) explore variants of LSTM to improve per-frame action classification whereas the method in Yeung et al. (2016) learns to directly predict action boundaries. An alternative approach based on 3D CNNs and temporal action proposals has shown competitive results in Shou et al. (2016) while Zhao et al. (2017) extends proposals in time and segments them in stages to evalute their “completeness” based on structured pyramid pooling. Here we use GRU as recurrent units for temporal modeling of actions in our RecLNet. Unlike previous work on temporal action localization, however, we address a much more challenging task of localizing actions in space and time.
Spatio-temporal detection. Spatio-temporal action detection aims to find locations of actions in space and time. The list of datasets for this task is limited: UCF101-24 – a subset of UCF-101 Soomro et al. (2012) used in the THUMOS challenge Gorban et al. (2015), the DALY dataset Weinzaepfel et al. (2016). Other datasets that we are aware of have either a very limited number of examples or consists of temporally trimmed videos. Some of the earlier works explore volumetric features and 3D sliding window detectors Ke et al. (2005); Laptev and Pérez (2007). More recent methods have extended ideas of object proposals in still images to action proposals in video Gkioxari and Malik (2015); Oneata et al. (2014); van Gemert et al. (2015). The common strategy in Gkioxari and Malik (2015); Peng and Schmid (2016); Saha et al. (2016); Singh et al. (2017); Weinzaepfel et al. (2015, 2016) is to localize actions in each frame with per-frame action or person detectors and to link resulting bounding boxes into continuous tracks. The temporal localization is then achieved by the temporal sliding windows Weinzaepfel et al. (2015, 2016) or dynamic programming Peng and Schmid (2016); Saha et al. (2016, 2017); Singh et al. (2017). Instead of relying on per-frame detections, Saha et al. (2017) regress pairs of successive frames and Hou et al. (2017) generate clip proposals from 3D feature maps. An approach of Zolfaghari et al. (2017) uses a CNN to sequentially fuse human pose features in addition to the standard appearance and optical flow modalities. Most recent method Kalogeiton et al. (2017) relies on SSD detector Liu et al. (2016b) adapted to spatio-temporal anchors and generates human tracks with tubelets linking. Soomro and Shah (2017); Yang and Yuan (2017) investigate unsupervised spatio-temporal action localization but this is outside of the scope of our work. In this paper, we propose a recurrent localization network (RecLNet) that both classifies and localizes actions within tracks supported by a thresholding and filtering method. Our analysis shows that RecLNet provides significant gains due to accurate temporal localization and its complementarity to the recent method Kalogeiton et al. (2017) with more accurate spatial localization but approximate temporal localization. Our resulting approach outperforms the state of the art in spatio-temporal action detection Kalogeiton et al. (2017); Peng and Schmid (2017); Saha et al. (2016); Singh et al. (2017); Weinzaepfel et al. (2016) on two challenging benchmarks for this task.
3 Action localization
This section presents our method for spatio-temporal action localization. The overview of the method is illustrated in Figure 2. A spatially localized person track (Figure 2, row 1) is passed to appearance and optical flow feature extractors (Figure 2, row 2). These descriptors feed our localization network composed of 3 layers for each stream and 1 fusion layer. In each stream, the first layer (Figure 2, row 3) normalizes either appearance or flow features before sending them to a stack of two GRU layer units. The GRU outputs from both stack levels and both streams are concatenated (Figure 2, row 5) and converted by a fully-connected layer (Figure 2, row 6) to action probabilities (Figure 2, last row).
In the following, we first briefly present the inputs of our method, namely the human tracks and their associated features. Then, we introduce our spatio-temporal action localization method based on the recurrent localization network (RecLNet). Finally, we discuss how to post-process the action detection scores in order to output final spatio-temporal human action tubes.
3.1 Person tracks
We obtain human tracks as in Saha et al. (2016); Weinzaepfel et al. (2016) by first running an action or person detector in each video frame, and then linking detections into human tracks spanning the entire video clip. Our method aims at segmenting the track in time to obtain temporal action boundaries. For this purpose, we associate to each time frame its corresponding bounding-box in the human track. This box is used as a pooling region to extract per-frame descriptors. Such features are obtained by Fast-RCNN Girshick (2015) appearance and flow ROI-pooling. The details on state-of-the-art human tracks and their associated features used in this work will be discussed in Section 4.
3.2 Temporal action localization
This section presents our model for action localization and its training procedure on the level of person tracks. In our work, we choose to adopt a model with memory links, like in a recurrent neural network (RNN), which we call recurrent localization network (RecLNet), to temporally localize actions. RNNs have been demonstrated to successfully model sequential data especially for language tasks such as speech recognition Dahl et al. (2012), machine translation Bahdanau et al. (2014) or image captioning Karpathy and Fei-Fei (2015). Given this success, we believe that recurrent networks are well-suited for modeling temporal sequences of appearance and motion in person tracks. Our final RecLNet model is composed of gated recurrent units (GRU) described below.
The LSTM and GRU architectures. Features from a human track of length can be seen as an input sequence where for each we aim to provide action activation , forming the output . To generate such output, we investigate two types of recurrent networks for our RecLNet, namely LSTM and GRU as defined below.
In the long short-term memory (LSTM) architectureHochreiter and Schmidhuber (1997)
, one memory cell and three ‘gates’ give LSTM the ability of discovering long-range temporal relationships by reducing the vanishing gradient problem compared to vanilla RNN. This is a useful property in our task since we need to handle particularly long human tracks. The LSTM cell takes as input featuresat time step together with the output at the previous time step and operates as follows:
is the sigmoid function and, , , are the ‘forget gate’, ‘input gate’, ‘output gate’ and ‘memory cell’, respectively. Matrices ,
and vectorsdenote the parameters of the cell.
The GRU Cho et al. (2014) cell differs from LSTM by the absence of the output gate and operates as follows:
where and are the ‘update gate’ and ‘reset gate’, respectively. The GRU cell is simpler, has less parameters and has shown some improvement on video tasks as in Tokmakov et al. (2017). An empirical comparison of LSTM and GRU cells is given in Section 5.
We define the Recurrent Localization Network (RecLNet) as a multi-class recurrent network trained to classify actions against background. As shown in Figure 2, at each time step, appearance (RGB) and optical flow (OF) Fast-RCNN networks take bounding boxes of human tracks as object proposals and extract features. Each stream (RGB and OF) is processed independently by feeding Fast-RCNN outputs (FC7 layers) to our RecLNet which produces action scores at each time step. Each stream of our RecLNet consists of a fully-connected layer that normalizes the input features (appearance or flow) and a stack of two GRU layers. Note the second GRU layer takes as input the output of the first one while their output is concatenated to a -dimensional stream output (where is the memory size). Finally the appearance and flow branch outputs are concatenated and a last fully-connected layer with softmax converts the recurrent output to an action probability for all actions (and background). The network is trained using the standard negative log-likelihood loss w.r.t. all action classes and background boxes:
with probabilities defined by softmax:
Here is the output of the last fully-connected layer corresponding to class , and denote features and labels at time step , symbol denotes network parameters, is the number of boxes in the training set, is the number of action classes, is the indicator function.
To train RecLNet, we set the ground-truth targets in the following way. We assign label to a frame bounding-box detection from a person track if it overlaps more than 0.3 spatial IoU with one ground truth annotation from action label , otherwise this input is considered as background. Having low IoU threshold allows us to get more positives.
Appearance and optical flow fusion. Single-stream networks are first independently trained then we consider three fusion methods to combine their appearance (RGB) and optical flow (OF) ouputs. The average simply averages both softmax RGB and OF network outputs. Both next fusion methods are trained on top of the two stream networks (their weights remaining fixed). The gating layer
learns per-class weights to multiply both network outputs before their original softmax layer and applies softmax after summing weighted class outputs. The third method,fusion layer, trains a fully-connected classification layer (followed by softmax) on top of the concatenated memory units outputs from each stream network (concatenation layer in Figure 2).
3.3 Post-processing of human action tracks
RecLNet output represents action scores at each time step of a human track. While being spatially localized, these scored tracks have to be segmented in time in order to produce the final spatio-temporal detections. For the track ranking, it is also necessary to score each of the final detections.
In this section, we then describe our method to get the final spatio-temporal detections and their associated score. Temporal localization is performed within each track.
Threshold. Let be the action to segment and a human track spanning from frame to of scores with the score associated to the person box at time . The goal of temporal segmentation is to extract from one or several sub-tracks (of time interval ) as final spatio-temporal detections.
For this purpose, we first get smoother scores by applying a median window filtering on the . Then, we temporally segment the track by selecting consecutive boxes with scores above a certain threshold , while others are rejected. More formally, considering that consecutive boxes from time to have already been added to the sub-track , the next box is added to if . Otherwise, ends and is returned as final detection. This method allows to break the initial track into several sub-track candidates of arbitrary lengths and is then able to capture several repetitive action instances happening on the same human track (like drinking, applying make up, playing harmonica, see Figure 4 and Figure 5). Here, model outputs must be smooth in order to get accurate action temporal boundaries. The recurrent units (like GRU, LSTM) are then well-suited for this localization method while appearance and flow CNNs output is generally noisier. Our temporal localization technique is referred to as threshold.
Temporally segmented track scoring. In order to rank detections and perform the average-precision (AP) evaluation, we need to set a score for each final spatio-temporal detection . Following, e.g., Saha et al. (2016), we define as the average of the top 40 action scores contained in . The same scores are also used for non-maximum-suppression (NMS) of spatio-temporal detection candidates based on their scores and overlap. For NMS, we use spatio-temporal Intersection-over-Union111The spatio-temporal IoU between two tracks is defined as a product of temporal IoU between the time segments of the tracks and average spatial IoU on the frames where both tracks are present. as overlap criterion.
3.4 Implementation details
RecLNet parameters. The FC7 output of each Fast-RCNN is -dimensional and the first fully-connected layers convert each stream to a -dimensional vector. The memory size is equal to 256 and the last fully-connected layer input is -dimensional (, i.e. the memory stack from both GRU layers of both streams).
Training. Both appearance and flow branches of RecLNet are separately trained using the Adam optimizer Kingma and Ba (2014) with a weight decay set to to avoid overfitting. Note that to train the single-stream networks, we halve the input dimension of the last fully-connect layer of RecLNet (concatenation layer in Figure 2). Training batches contain different tracks of temporal length
. Backpropagation through time (BPTT) is then performed everytime steps.
Detection. In all experiments, the NMS overlap threshold is set 0.2, the action localization threshold is set to and the median window size is 25.
Optical flow. To obtain the optical flow data, we compute horizontal and vertical flow for each consecutive pair of frames using the approach of Brox et al. (2004). Following Gkioxari and Malik (2015); Weinzaepfel et al. (2015), flow maps are saved as 3-channels images corresponding to optical flow in x and y direction and its magnitude with all the values restricted to the interval .
4 Experimental setup
This section describes UCF101-24 Soomro et al. (2012) and DALY Weinzaepfel et al. (2016) datasets used for evaluation of our method. For both datasets, we provide experimental details on the use of ground truth annotation and data pre-processing.
The original version of the UCF-101 dataset Soomro et al. (2012) is designed for action classification and contains 13321 videos for 101 action classes. The task of spatio-temporal action localization is defined on a subset of 24 action classes (selected by Gorban et al. (2015)) in 3207 videos. We refer to this subset as “UCF101-24”. Each instance of an action is manually annotated by a person track with the temporal interval corresponding to the interval of an action. In our training and testing, we use recently corrected ground truth tracks Saha et al. (2016).222https://github.com/gurkirt/corrected-UCF101-Annots Each UCF-101 video contains actions of a single class.
Short vs. long classes. While some of the UCF101-24 action classes are short (’basketball dunk’ or ’tennis swing’) other continuous actions (’biking’ or ’rope climbing’) typically last for the full duration of the video. To better evaluate the temporal localization on UCF101-24, we define a subset with short action classes that on average last less than a half of the video length. This subset contains six actions (Basketball, Basketball Dunk, Cricket Bowling, Salsa Spin, Tennis Swing and Volleyball Spiking) and we call them “Short classes”. In Section 5, we evaluate localization for all 24 action classes and for Short classes separately.
Performance evaluation. To evaluate the detection performance, we use the standard spatio-temporal IoU criterion defined by the UCF101-24 benchmark. The detected action tube is considered to be correct if and only if its intersection with the ground-truth tube is above the criterion threshold and if both tubes belong to the same action class. Duplicate detections are considered as false positives and are penalized by the standard precision-recall measure. The overall performance on the UCF101-24 dataset is compared in terms of mean average precision (mAP) value.
Action-specific human tracks. To enable direct comparison of our method with Saha et al. (2016), for UCF101-24 experiments we use the same person tracks as in Saha et al. (2016). These tracks are obtained by linking per-frame action detections with dynamic programming (DP). Action detections are obtained with the Fast-RCNN method Girshick (2015) trained with appearance and flow input for the task of spatial action localization. As our model performs its own temporal localization, we do not run the temporal segmentation of Saha et al. (2016) (second DP pass) and keep 5 action proposals per action covering the whole video. Per-frame input features for our RecLNet are obtained from the same Fast-RCNN detector used for track detection.
DALY Weinzaepfel et al. (2016) is a recent large-scale dataset for action localization containing 510 videos (31 hours) of 10 different daily activities such as ’brushing teeth’, ’drinking’ or ’cleaning windows’. There is only one split containing 31 train and 20 test videos per class. The average length of the videos is 3min 45s. In contrast to UCF101-24, all actions are short w.r.t. the full video length, making the task of temporal action localization more challenging. DALY may contain multiple action classes in the same video. The DALY dataset provides ground-truth temporal boundaries for all the action instances and the spatial annotation (bounding boxes) for few keyframes of each instance.
Human tracks and associated features. To enable direct comparison with Weinzaepfel et al. (2016) we use tracks provided by Weinzaepfel et al. (2016). An accurate Faster-RCNN Ren et al. (2015) person detector is trained on the large MPII human pose dataset Andriluka et al. (2014). The tracks are obtained by linking human detections with a tracking-by-detection approach. Similarly to UCF101-24, the appearance and flow features are obtained with a Fast-RCNN detector trained to detect actions on the annotated DALY frames. The action scores from this detector will be used as a baseline in Section 5 and referred to as CNN RGB+OF. Following Mathias et al. (2014)
, we compensate the annotation bias by adapting human box sizes to action annotation. To achieve this, we train a linear regression of bounding boxes. For each DALY annotated keyframe from the training set we associate its most overlapping bounding-box from the human tracks (with at least 0.5 IoU) to train the linear model.
DALY tracks labeling. To compensate for the sparse action annotation in DALY, we extend ground truth to all frames of the actions with automatic tracking. We use the Siamese Fully-Convolutional Network for online non-class-specific tracking method333https://github.com/bertinetto/siamese-fc Bertinetto et al. (2016) initialized on all ground-truth keyframes. For a given action instance, we aggregate all tracks at each frame by a median bounding box. To assess the quality of generated ground-truth tracks, we computed the track proposals recall ( at ) within the action time interval. Also, the IoU overlap between ground-truth tracks and annotated keyframes is greater than (resp. ) for (resp. ) of keyframes.
In this section, we first evaluate the impact of the threshold temporal localization, recurrent architecture and the fusion method (Section 5.1). We then evaluate the potential gain due to action classification (Section 5.2) followed by an extensive analysis on temporal localization (Section 5.3). Next, we show an improvement if using I3D and temporal person tracks (Section 5.4). We compare our RecLNet to the state of the art on UCF101-24 and DALY datasets (Section 5.5). Section 5.6 concludes this experimental part by presenting qualitative results.
5.1 Impact of localization method, recurrent architecture and fusion
|CNN Saha et al. (2016)||59.3||55.2||60.4||-||-||-|
In Saha et al. (2016), temporal localization is performed using the Viterbi algorithm on top of action tracks spanning the whole video. To temporally trim these tracks, one binary label (action vs. background) is associated to each bounding-box by maximizing an energy function where unary potentials represent box action scores while pairwise potentials control the final action length (or track smoothness). This smoothness score is weighted per class and then might over control the track length towards the action durations seen on the training set while, in our method, median filtering encodes smoothness more explicitly and suffers less from the dataset biases described in 4.1.
In Table 2, we refer to the track action scores from Saha et al. (2016) as CNN RGB+OF, since they come from two Fast-RCNN where appearance and optical flow outputs have been fused. We recall that to enable a direct comparison with Saha et al. (2016) we are using the same human tracks as Saha et al. (2016) but score them differently (see Section 4 for details). Here, we compare the Viterbi algorithm originally used in Saha et al. (2016) for temporal localization and our threshold method on original scores (CNN RGB+OF). Interestingly we observe that our threshold technique improves over their original results by and conclude that thresholding combined with median filtering works better than this typically used localization method while being simpler. The next paragraph shows the impact of differently scoring, among others, tracks from Saha et al. (2016). In the following, only our threshold temporal localization is used.
|CNN RGB+OF||Viterbi Saha et al. (2016)||55.5|
Recurrent architecture. Table 1 compares models based on LSTM and GRU units (see Section 3 for details). As a baseline we use a model with no temporal connections but similar architecture: the recurrent units are replaced by a stack of 2 fully-connected layers with non-linearity and has the same number of parameters as the GRU unit. This additional fully-connected classifier is referred to as FC. For UCF101-24 evaluation, we again report original track scores (CNN Saha et al. (2016) as in Table 2) while for DALY, CNN represents the Fast-RCNN scores similarly retrained as authors Weinzaepfel et al. (2016) on their human tracks (see Section 4.2 for details). Input modalities are either, optical flow features (OF), appearance features (RGB) or both (RGB+OF) where action scores from the two stream outputs are averaged in this case. Spatio-temporal action localization is evaluated on UCF101-24 and DALY datasets at spatio-temporal IoU of 0.3.
We first observe that training an additional classifier (FC) is better than directly taking Fast-RCNN outputs (CNN) and improves performance by and on UCF101-24 and DALY respectively when both appearance and optical flow streams are used (RGB+OF). Also, non-recurrent baselines (CNN and FC) perform worse than the recurrent variants (LSTM and GRU). When using both RGB+OF as input, the gain due to recurrence is and on UCF101-24 and DALY respectively when comparing GRU to CNN and and when comparing GRU to FC. Interestingly, recurrent unit improvement is larger on optical flow on UCF101-24 and on DALY when comparing to . This shows that temporal memory links are beneficial for better spatio-temporal action localization performance. We also note, as this is often the case when working with videos (e.g. Tokmakov et al. (2017)), that GRU outperforms LSTM while being a simpler model.
This experiment has shown GRU memory unit achieves the best action localization accuracy. In the following, GRU will then be used in our recurrent model for determining the action temporal extent in all experiments.
Fusion method. Table 3 explores the different fusion strategies described in Section 3 to combine appearance (RGB) and flow (OF) features in our localization model (with GRU layer which was validated in the previous experiment). We first note that all fusion methods improve action localization results on both UCF101-24 and DALY. However, the simple averaging method is not able to capture feature complementarity between appearance and flow features especially on UCF101-24 where it gets only improvement compared to the best performing OF features. The gating layer is able to take advantage of features combination by substantially improving on both datasets. Finally, the fusion layer better captures features complementarity and improves action localization mAP on UCF101-24 and DALY respectively by and compared to the single flow features and by and compared to appearance features.
In the following, our final recurrent model uses the fusion layer and is referred to as RecLNet in all experiments.
Finally, when using our threshold method, re-scoring the tracks with our RecLNet instead of taking Fast-RCNN outputs (CNN) as in Saha et al. (2016); Weinzaepfel et al. (2016) improves spatio-temporal action localization results by and on UCF101-24 and DALY respectively (from Table 1). The improvement is even larger () when comparing RecLNet accuracy () to the original result using scores and temporal localization from Saha et al. (2016) ( on UCF101-24 from Table 2).
Given that we are using the same tracks, this first gives an insight that our method improvement compared to Saha et al. (2016); Weinzaepfel et al. (2016) might be due to temporal localization. This question is further studied in the next sections.
5.2 Analyzing the performance gain: action classification
Spatio-temporal action localization is composed of spatial localization, action classification and temporal localization. In this section we focus on action classification and analyze its performance given person tracks and pre-defined temporal action boundaries.
Evaluating action classification. Here, we do not require temporal localization by restricting the track to the ground-truth time interval in order to evaluate the performance gain due to action classification. As described in Section 3, the track scoring is obtained with the average over the top 40 action scores in the trimmed interval. Table 4 shows our method (RecLNet) increases action classification by approximately compared to scores from Saha et al. (2016) (CNN RGB+OF) when evaluated on UCF101-24 classes. This improvement is moderate compared to the spatio-temporal localization boost of Section 5.1.
|CNN RGB+OF Saha et al. (2016)||84.3||81.6||75.6|
Table 5 shows the same experiments on DALY. We observe that our RecLNet method performs slightly worse (from to ) compared to the CNN RGB+OF baseline (see Section 4.2 for details). This might be due to DALY track labeling which, contrary to UCF101-24, is obtained by automatic tracking of ground-truth sparse annotations (see Section 4.2) which introduces some noise and can slightly affect action classification. We, therefore, conclude that the spatio-temporal action localization gain of our method over the baseline, demonstrated in Section 5.1, cannot come from action classification.
Overall, this experiment confirms results in the literature Ng et al. (2015) that recurrent units (RNN, LSTM, GRU) do not really improve the performance for action classification. The spatial localization (the person boxes positions) being fixed, the gain of our method for spatio-temporal action localization has to come from better temporal localization. This observation is analyzed in detail in Section 5.3.
|Saha et al. (2016)||9.8||11.9||7.3||18.8||1.0||27.2||12.7||56.1||67.1||82.9||80.5||99.7||59.8||88.1||64.1||40.1||49.6||78.4||85.1||70.7||92.7||84.9||55.4||36.9||67.1||55.5|
|Kalogeiton et al. (2017)||18.0||44.1||23.5||20.7||1.0||39.1||24.4||72.0||82.0||84.8||81.3||99.2||76.7||94.6||60.8||82.5||90.4||92.4||92.0||84.0||80.4||91.4||65.9||62.3||76.8||67.3|
5.3 Analyzing the performance gain: temporal localization
In this section, we show the main gain of RecLNet comes from better temporal localization and demonstrate its complementarity with the state of the art Kalogeiton et al. (2017). We first motivate this study by explaining the UCF101-24 bias, then a per-class analysis compares our RecLNet to Kalogeiton et al. (2017) and to the method we build on Saha et al. (2016). Finally, we investigate the potential room for improvement.
UCF101-24 bias. As described in Section 4.1, UCF101-24 contains only six short action classes lasting less than half of the video duration, while 17 actions ( of the dataset classes) span at least of the video length in which 11 of them ( of the dataset classes) span even more than of the video length. UCF101-24 results are then biased toward long actions which do not require temporal localization. Indeed, tracks spanning the whole video already achieves good temporal localization for these classes (note that such detections would totally fail on the DALY dataset). To avoid this bias and to focus on challenging cases of temporal localization, we next compare our method to Kalogeiton et al. (2017); Saha et al. (2016) on the subset of short classes along with whole dataset.
Per-class performance on UCF101-24. Table 6 compares the per-class results of our approach with Saha et al. (2016) since we are using their human tracks and with the best performing state-of-the-art method Kalogeiton et al. (2017) (see Section 5.5 for comparison with the state of the art). We report mAP for short classes only (mAP-short) and also for all classes (mAP). We outperform Saha et al. (2016) by on average on short classes and by on “Basketball”. As we have already seen previously, the overall improvement is . We can observe that for some “long” actions the performance drops slightly. Similarly, we outperform the state of the art Kalogeiton et al. (2017) by only overall while we reach a large improvement of around on short classes.
|All classes||Short classes|
|Kalogeiton et al. (2017)||67.3||51.4||22.7||24.4||2.6||0.0|
Table 7 now compares our model to the best performing method Kalogeiton et al. (2017) at higher IoU. We observe that even if short classes get extremely difficult to detect at highest IoU ( mAP at IoU 0.75), at IoU 0.5, our model still outperforms the state of the art by on short classes while loosing overall (Section 5.5 studies this latter result when comparing to the state of the art).
These experiences on per-class performance show that our model is able to improve Saha et al. (2016) by producing more accurate temporal action localization. Also, while Kalogeiton et al. (2017) takes advantage from strong features and spatially accurate person boxes to get excellent accuracy on long classes (e.g. compared to Saha et al. (2016)), its performance on short classes suffers from approximate temporal localization. This demonstrates the potential complementarity between our RecLNet and the current best performing method Kalogeiton et al. (2017). The room for improvement of RecLNet will be studied in the next section.
|Correct||Correct||Correct||class mAP @0.75|
|temporal loc.||spatial loc.||class||short||all|
Toward action localization improvement. This paragraph analyzes how to improve the spatio-temporal action localization performance of our model especially when evaluating at very high IoU (0.75). To distinguish the potential improvements, we can consider action classification, spatial localization and/or temporal localization as being correct. We proceed as follow. When we suppose perfect action classification, we set final spatio-temporal detection score (see Section 3.3) to 0 for all false positives (note this is equivalent to the recall). Let (resp. ) the spatial IoU overlap of a final detection with its ground-truth. Thereby, when considering perfect spatial (resp. temporal) localization, if (resp. ) is greater than 0.3, it is then set to 1 in the spatio-temporal IoU computation. We choose 0.3 as it seems fair enough 1) not to get a track completely shifted in time () for which the spatial IoU would have been computed on only one frame or a few and 2) to ensure that the track at least approximately spatially “follows” () the person. Table 8 presents the combinations of these assumptions. First, assuming correct temporal localization or action classification only improves our model performance by respectively and . However, the spatial assumption is by far the best room for improvement of our model () and achieves accuracy. It also shows our method gets accuracy on short classes mostly because of inaccurate human track bounding-boxes as the spatial assumption improves it to . Of course, combining several assumptions further improve the performance (note that last line does not reach since the track recall is not and that the above “0.3 criterion” eliminates some track candidates). However, we observe that combining correct temporal localization and action classification () is still far from the single spatial localization assumption (). This experiment confirms that improving spatial human detection is the direction to take further enhancements and demonstrates that temporal localization is already a strong component of RecLNet.
This section has shown that the UCF101-24 bias is not in favour of our model while the short classes are by far the hardest to localize and then are source of potentially large improvement for current action localization methods. Also, we observed that our model already benefits from accurate temporal localization while its considerable room for improvement is the spatial localization component (the temporal localization being the one with the less potential for RecLNet). Consequently, the state of the art Kalogeiton et al. (2017), which mostly relies on spatially more accurate human tracks and better features that the ones we build on Saha et al. (2016) (as previously compared in Table 8), and our model, which greatly improves temporal localization, are definitely complementary.
5.4 Improved tracks and descriptors
By using the same tracks and features as input, the previous sections have shown our RecLNet is able to correct the temporal localization weakness of state-of-the-art methods. Here, we substitute these tracks and features (introduced in Section 3.1 and 4) by improved models. In the following, we first explain the track modification then the features substitution and finally analyze their impact.
Person action tracks with temporal integration. In order to improve person tracking, we integrate temporal information in the detector. Supported by our analysis in Section 5.3, showing that better spatial localization would greatly improve our results, and by Kalogeiton et al. (2017) that shows large detection improvement by stacking features coming from several neighboring frames, we design a new approach for tracks. In the same spirit as Kalogeiton et al. (2017) who stacks frame features in the SSD detector Liu et al. (2016b), we adapt Faster-RCNN Ren et al. (2015) to perform accurate detection by temporally integrating stacks of images. We introduce the following modifications to the Faster-RCNN pipeline. First, the inference pass, producing the feature map on which the ROI-pooling applies, is performed independently on consecutive frames. The output feature maps are stacked along the channel dimension and will serve at computing scores and regressions. Second, the RPN computes at each anchor location a global objectness score for the stack and regressions. Third, anchors labels are computed based on the mean overlap between the ground-truth boxes and the regressed proposals. Fourth, the 3D proposals are further regressed to action proposals (tubelets) by computing one action score and regressions per class. Tubelets are linked into action tracks using the code of Kalogeiton et al. (2017). This is an online method that iteratively aggregates tubelets sorted by action score to a set of current links based on spatio-temporal overlap. The detector is used on the ResNet-101 architecture He et al. (2016) and we choose since it is a good trade-off between tubelet quality and efficiency. Since only sparse keyframes are spatially annotated on DALY, automatic tracking is performed to propagate the ground truth (see Section 4.2). Here, stacks used at train time on DALY are then composed of one annotated keyframe and its 2 frames tracked forward and backward. These action tracks are referred to as stack in the following.
I3D features. Recent results have shown large action recognition improvement using the I3D Carreira and Zisserman (2017b) architecture. We replace the per-frame features by descriptors extracted with the I3D RGB and flow networks both trained on the Kinetics dataset Kay et al. (2017b)
. We extract features after the 7-th inception block, before the max-pooling, as it is a good balance between deepness, for strong classification results, and accurate resolution, for precise track pooling. After this block, the temporal receptive field is roughly one hundred frames. As input, we use a spatial resolution ofpixels resulting in feature maps of size with channels. These feature maps are extracted at intervals of 4 frames. The box temporally aligned with the middle of the interval is used as pooling region to obtain a -dimensional I3D descriptor representing the track at this time step.
|RecLNet||f. based||f. based||69.0||57.5||46.5||30.2||25.4||19.7|
Experimental evaluation. Here, frame-based features and frame-based tracks we refer to are the ones described in Section 4. Table 9 first shows that when substituting the baseline features by the I3D descriptors our RecLNet model obtains a performance boost of around on UCF101-24 ( at IoU 0.5) and - on DALY ( at IoU 0.3). This result is inline with Gu et al. (2017). When we further replace the frame-based tracks by ours obtained with stack-Faster-RCNN (stack), RecLNet results get a second improvement of - on UCF101-24 and - on DALY. This gain of performance validates what was shown by Kalogeiton et al. (2017), i.e, integrating temporal information in the detector increases the track spatial precision leading to action localization improvement (as we observed in Section 5.3). For comparison we also report the fully connected classifier (FC) results as in Section 5.1. Given the large temporal receptive field of I3D features mentioned earlier, we notice RecLNet still improves the performance, e.g, it gets and mAP at IoU 0.3 when using the stack tracks on UCF101-24 and DALY respectively. In the following, the RecLNet model with the I3D features and the improved tracks (stack) is referred to as RecLNet++.
5.5 Comparison to the state of the art
|Weinzaepfel et al. (2015)||54.3||51.7||46.8||37.8||-||-||-||-|
|Hou et al. (2017)||54.7||51.3||47.1||39.2||-||-||-||-|
|Zolfaghari et al. (2017)||65.2||59.5||47.6||38.0||-||-||-||-|
|Peng and Schmid (2016)||54.5||50.4||42.3||32.7||-||-||-||-|
|Weinzaepfel et al. (2016)||71.1||-||58.9||-||-||-||-||-|
|Peng and Schmid (2017)||78.8||77.3||72.9||65.7||-||-||-||-|
|Saha et al. (2016)||79.1||76.6||66.8||55.5||46.4||35.9||26.8||-|
|Mettes et al. (2016)||-||-||34.8||-||-||-||-||-|
|Singh et al. (2017)||-||-||73.5||-||-||46.3||-||15.0|
|Saha et al. (2017)||-||71.3||63.1||51.6||-||33.1||-||-|
|Kalogeiton et al. (2017)||-||-||76.5||65.2||-||49.2||-||19.7|
|Gu et al. (2017)||-||-||-||-||-||59.9||-||-|
UCF101-24. Table 10 compares RecLNet and RecLNet++ models to Gu et al. (2017); Hou et al. (2017); Kalogeiton et al. (2017); Mettes et al. (2016); Peng and Schmid (2016, 2017); Saha et al. (2016, 2017); Singh et al. (2017); Weinzaepfel et al. (2015, 2016); Zolfaghari et al. (2017) on UCF101-24 spatio-temporal action localization. RecLNet is our recurrent network with GRU memory units and fusion layer to combine optical flow and appearance features performing localization with our threshold approach and RecLNet++ uses I3D features and our improved stack tracks. RecLNet already significantly outperforms all other methods for all IoU thresholds below . Due to our significantly improved temporal localization, our approach outperforms Saha et al. (2016) on which we build by at high IoU (). We also outperform Peng and Schmid (2017), despite the fact that they use more sophisticated features combining a number of different human parts as well as multi-scale training and testing. Both these components are complementary to our approach. Our significant boost in performance can be explained by the fact that most current action localization methods rely on spatial features, for example per-frame CNN descriptors for optical flow and appearance, to detect action both spatially and temporally. However, such features are not designed for temporal localization. By relying on an accurate temporal model such as our recurrent RecLNet, we can clearly improve temporal detection and methods we build on that do not focus on this aspect. Recent method Kalogeiton et al. (2017), relying on more accurate spatial localization obtained by spatio-temporal cuboid regression, outperforms our model at high IoU. Indeed, they benefit from more precise human track bounding-boxes than the one we build on. Such an approach is used by RecLNet++, the enhanced version of RecLNet. This is shown to improve the state of the art by a large margin outperforming Kalogeiton et al. (2017) by , and at IoU 0.3, 0.5 and 0.75 respectively. We also mention very recent work Gu et al. (2017) which only reports results at IoU 0.5 and achieves similar performance.
|Weinzaepfel et al. (2016)||sliding window||-||14.5||-|
|Action classes||CNN RGB+OF||RecLNet|
|App. MakeUp Lips||3.4||13.6|
DALY. The recent DALY dataset Weinzaepfel et al. (2016) is very challenging for spatio-temporal localization because it contains only short actions compared to the video length making temporal segmentation crucial and difficult. Table 11 compares our results to the state of the art. We can see that our RecLNet approach significantly outperformst Weinzaepfel et al. (2016) by more than . We also report results with our CNN scores CNN RGB+OF and temporal thresholding of the action scores without using RecLNet. Interestingly, the baseline approach outperforms the state of the art, but performs significantly worse than our (RecLNet) method. Also, Table 12 shows RecLNet improves over the baseline for all classes except Drinking. Again, since we build on tracks from Weinzaepfel et al. (2016) and extract similar features (see Section 3), the improvement should be attributed to the more accurate temporal localization achieved by our RecLNet model. RecLNet++ further improves the results and outperforms Weinzaepfel et al. (2016) by .
5.6 Qualitative results
Good temporal localization needs precise evolution of action score with clear temporal boundaries. Having smooth scores also help our threshold localization method presented in Section 3.3. Here, we provide qualitative results to illustrate that our RecLNet possesses these properties and is well-suited for such localization.
Figure 3 shows scores from Saha et al. (2016) (red curves) versus our RecLNet response (green curves) for the 5 action track proposals used on UCF101-24. Thus, there are in total 10 curves per graph, one curve per track for the two compared methods. A bold curve (resp. dashed curve) shows that its corresponding track overlaps with more (resp. less) than spatial IoU with a ground-truth action annotation. The ground-truth action interval is represented by the horizontal blue bar below the x-axis. The x-axis represents time. For each graph, we show the frame corresponding to the maximum RecLNet response (localized by the vertical yellow line) with its associated human detection (yellow box).
We show scores for 3 short actions basketball, tennis swing and volleyball spiking, and 2 long actions, cliff diving and diving. We can observe that the scores of the RecLNet correspond to the temporal boundaries more precisely than the CNN scores of Saha et al. (2016). Scores from Saha et al. (2016) are often flat and have high values throughout the entire video length (Figure 3, row 3). In general, they are not precise temporally (e.g., Figure 3, row 4). Also, a lot of dashed lines are highly scored (Figure 3, rows 3 and 5) which adds false positives. Finally, these graphs show that it is easy to set a threshold on RecLNet responses in order to obtain good temporal localization, while scores from Saha et al. (2016) (and scores from CNNs in general) are often imprecise temporally.
Similarly, Figures 4, 5 show qualitative results for temporal action localization on the DALY dataset, comparing again our RecLNet (green curves) and the CNN RGB+OF baseline (red curves). Each curve corresponds to one person track in the video. Each row corresponds to one action class. The last (third) column in Figures 4, 5 corresponds to failure cases or examples where our RecLNet model does not improve temporal localization. Similarly to results for the UCF101-24 dataset, the two first columns shows that scores of RecLNet are better aligned with the ground truth action boundaries compared to scores of the CNN RGB+OF baseline. At the same time, dashed lines indicate many high-scoring false positive detections for the CNN RGB+OF method (e.g. Figure 4, row 1, plot 2 and Figure 5, row 2, plot 2). The last column in Figures 4, 5 indicates that DALY is a challenging dataset with many difficult examples where, most of the time, both methods fail to detect the correct temporal action boundaries.
This paper shows that training a recurrent model (GRU) on the level of person tracks for modeling the temporal structure of actions improves localization in time significantly. Building on current state-of-the-art methods that obtain very good results for spatial localization, RecLNet improves significantly the detection of action time boundaries and hence the overall performance of spatio-temporal localization. As demonstrated in our analysis, improving spatial precision of human tracks, inspired by most recent methods, and using better features further improve our results (RecLNet++). Our method outperforms the state of the art on the two challenging datasets, namely UCF101-24 and DALY.
This work was supported in part by ERC grants ACTIVIA and ALLEGRO, the MSR-Inria joint lab, the Louis Vuitton ENS Chair on Artificial Intelligence, an Amazon academic research award, and an Intel gift.
Andriluka et al. 
M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele.
2d human pose estimation: New benchmark and state of the art analysis.In CVPR, 2014.
Baccouche et al. 
M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and A. Baskurt.
Sequential deep learning for human action recognition.In International Workshop on Human Behavior Understanding, 2011.
Bagautdinov et al. 
T. Bagautdinov, A. Alahi, F. Fleuret, P. Fua, and S. Savarese.
Social scene understanding: End-to-end multi-person action localization and collective activity recognition.In CVPR, 2017.
- Bahdanau et al.  D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align and translate. In CoRR, 2014.
- Bertinetto et al.  L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. Torr. Fully-convolutional siamese networks for object tracking. In BMVC, 2016.
- Brox et al.  T. Brox, A. Bruhn, N. Papenberg, and J. Weickert. High accuracy optical flow estimation based on a theory for warping. In ECCV, 2004.
- Carreira and Zisserman [2017a] J. Carreira and A. Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017a.
- Carreira and Zisserman [2017b] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017b.
- Cho et al.  K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. CoRR, 2014.
- Dahl et al.  G. E. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition. In IEEE Transactions on Audio, Speech, and Language Processing, 2012.
- de Souza et al.  C. R. de Souza, A. Gaidon, E. Vig, and A. M. López. Sympathy for the details: dense trajectories and hybrid classification architectures for action recognition. In ECCV, 2016.
- Donahue et al.  J. Donahue, L. A. Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, K. Saenko, and T. Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015.
- Feichtenhofer et al. [2016a] C. Feichtenhofer, A. Pinz, and R. P. Wildes. Spatiotemporal residual networks for video action recognition. In NIPS, 2016a.
- Feichtenhofer et al. [2016b] C. Feichtenhofer, A. Pinz, and A. Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016b.
- Girshick  R. Girshick. Fast r-cnn. In ICCV, 2015.
- Gkioxari and Malik  G. Gkioxari and J. Malik. Finding action tubes. In CVPR, 2015.
- Gorban et al.  A. Gorban, H. Idrees, Y.-G. Jiang, A. Roshan Zamir, I. Laptev, M. Shah, and R. Sukthankar. THUMOS challenge: Action recognition with a large number of classes. http://www.thumos.info/, 2015.
- Gu et al.  C. Gu, C. Sun, S. Vijayanarasimhan, C. Pantofaru, D. A. Ross, G. Toderici, Y. Li, S. Ricco, R. Sukthankar, C. Schmid, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CoRR, 2017.
- He et al.  K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016.
- Heilbron et al.  F. C. Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015.
- Hoai et al.  M. Hoai, Z.-Z. Lan, and F. De la Torre. Joint segmentation and classification of human actions in video. In CVPR, 2011.
- Hochreiter and Schmidhuber  S. Hochreiter and J. Schmidhuber. Long short-term memory. In Neural computation, 1997.
Hou et al. 
R. Hou, C. Chen, and M. Shah.
Tube convolutional neural network (t-cnn) for action detection in videos.ICCV, 2017.
- Idrees et al.  H. Idrees, A. R. Zamir, Y.-G. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah. The THUMOS challenge on action recognition for videos “in the wild”. In Computer Vision and Image Understanding, 2016.
- Ji et al.  S. Ji, W. Xu, M. Yang, and K. Yu. 3D convolutional neural networks for human action recognition. In ICML, 2010.
- Kalogeiton et al.  V. Kalogeiton, P. Weinzaepfel, V. Ferrari, and C. Schmid. Action tubelet detector for spatio-temporal action localization. In ICCV, 2017.
- Karpathy and Fei-Fei  A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015.
- Karpathy et al.  A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014.
- Kay et al. [2017a] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, M. Suleyman, and A. Zisserman. The kinetics human action video dataset. CoRR, 2017a.
- Kay et al. [2017b] W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, et al. The kinetics human action video dataset. In CoRR, 2017b.
- Ke et al.  Y. Ke, R. Sukthankar, and M. Hebert. Efficient visual event detection using volumetric features. In ICCV, 2005.
- Kingma and Ba  D. Kingma and J. Ba. Adam: A method for stochastic optimization. CoRR, 2014.
- Kuehne et al.  H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. HMDB: a large video database for human motion recognition. In ICCV, 2011.
- Laptev and Pérez  I. Laptev and P. Pérez. Retrieving actions in movies. In ICCV, 2007.
- Laptev et al.  I. Laptev, M. Marszałek, C. Schmid, and B. Rozenfeld. Learning realistic human actions from movies. In CVPR, 2008.
- Liu et al. [2016a] J. Liu, A. Shahroudy, D. Xu, and G. Wang. Spatio-temporal LSTM with trust gates for 3D human action recognition. In ECCV, 2016a.
- Liu et al. [2016b] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg. Ssd: Single shot multibox detector. In ECCV, 2016b.
- Ma et al.  S. Ma, L. Sigal, and S. Sclaroff. Learning activity progression in LSTMs for activity detection and early detection. In CVPR, 2016.
- Mathias et al.  M. Mathias, R. Benenson, M. Pedersoli, and L. Van Gool. Face detection without bells and whistles. In ECCV, 2014.
- Mettes et al.  P. Mettes, J. C. van Gemert, and C. G. M. Snoek. Spot on: Action localization from pointly-supervised proposals. In ECCV, 2016.
- Ng et al.  J. Y.-H. Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals, R. Monga, and G. Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015.
- Oneata et al.  D. Oneata, J. Revaud, J. Verbeek, and C. Schmid. Spatio-temporal object detection proposals. In ECCV, 2014.
- Peng and Schmid  X. Peng and C. Schmid. Multi-region two-stream R-CNN for action detection. In ECCV, 2016.
- Peng and Schmid  X. Peng and C. Schmid. Multi-region two-stream R-CNN for action detection. In HAL, 2017.
- Pigou et al.  L. Pigou, A. van den Oord, S. Dieleman, M. Van Herreweghe, and J. Dambre. Beyond temporal pooling: Recurrence and temporal convolutions for gesture recognition in video. In IJCV, 2017.
- Pirsiavash and Ramanan  H. Pirsiavash and D. Ramanan. Parsing videos of actions with segmental grammars. In CVPR, 2014.
- Ren et al.  S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NIPS, 2015.
- Rohrbach et al.  M. Rohrbach, A. Rohrbach, M. Regneri, S. Amin, M. Andriluka, M. Pinkal, and B. Schiele. Recognizing fine-grained and composite activities using hand-centric features and script data. In IJCV, 2016.
- Saha et al.  S. Saha, G. Singh, M. Sapienza, P. H. Torr, and F. Cuzzolin. Deep learning for detecting multiple space-time action tubes in videos. In BMVC, 2016.
- Saha et al.  S. Saha, G. Singh, and F. Cuzzolin. Amtnet: Action-micro-tube regression by end-to-end trainable deep architecture. ICCV, 2017.
- Shou et al.  Z. Shou, D. Wang, and S.-F. Chang. Temporal action localization in untrimmed videos via multi-stage CNNs. In CVPR, 2016.
- Simonyan and Zisserman  K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos. In NIPS, 2014.
- Singh et al.  B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for fine-grained action detection. In CVPR, 2016.
- Singh et al.  G. Singh, S. Saha, M. Sapienza, P. Torr, and F. Cuzzolin. Online real time multiple spatiotemporal action localisation and prediction. In CoRR, 2017.
- Soomro and Shah  K. Soomro and M. Shah. Unsupervised action discovery and localization in videos. In ICCV, 2017.
- Soomro et al.  K. Soomro, A. R. Zamir, and M. Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. In CoRR, 2012.
- Srivastava et al.  N. Srivastava, E. Mansimov, and R. Salakhutdinov. Unsupervised learning of video representations using LSTMs. In CoRR, 2016.
- Sutskever et al.  I. Sutskever, J. Martens, and G. E. Hinton. Generating text with recurrent neural networks. In ICML, 2011.
- Taylor et al.  G. W. Taylor, R. Fergus, Y. LeCun, and C. Bregler. Convolutional learning of spatio-temporal features. In ECCV, 2010.
- Tokmakov et al.  P. Tokmakov, K. Alahari, and C. Schmid. Learning video object segmentation with visual memory. In ICCV, 2017.
- Tran et al.  D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3D convolutional networks. In ICCV, 2015.
- van Gemert et al.  J. C. van Gemert, M. Jain, E. Gati, and C. G. M. Snoek. APT: Action localization proposals from dense trajectories. In BMVC, 2015.
- Varol et al.  G. Varol, I. Laptev, and C. Schmid. Long-term temporal convolutions for action recognition. PAMI, 2017.
- Vinyals et al.  O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In CVPR, 2015.
- Wang and Schmid  H. Wang and C. Schmid. Action recognition with improved trajectories. In ICCV, 2013.
- Weinzaepfel et al.  P. Weinzaepfel, Z. Harchaoui, and C. Schmid. Learning to track for spatio-temporal action localization. In ICCV, 2015.
- Weinzaepfel et al.  P. Weinzaepfel, X. Martin, and C. Schmid. Human action localization with sparse spatial supervision. In CoRR, 2016.
- Yang and Yuan  J. Yang and J. Yuan. Common action discovery and localization in unconstrained videos. In ICCV, 2017.
- Yeung et al.  S. Yeung, O. Russakovsky, G. Mori, and L. Fei-Fei. End-to-end learning of action detection from frame glimpses in videos. In CVPR, 2016.
- Yuan et al.  J. Yuan, B. Ni, X. Yang, and A. A.Kassim. Temporal action localization with pyramid of score distribution features. In CVPR, 2016.
- Zhao et al.  Y. Zhao, Y. Xiong, L. Wang, Z. Wu, X. Tang, and D. Lin. Temporal action detection with structured segment networks. ICCV, 2017.
- Zolfaghari et al.  M. Zolfaghari, G. L. Oliveira, N. Sedaghat, and T. Brox. Chained multi-stream networks exploiting pose, motion, and appearance for action classification and detection. ICCV, 2017.