We propose a novel approach based on deep Convolutional Neural Networks (CNN) to recognize human actions in still images by predicting the future motion, and detecting the shape and location of the salient parts of the image. We make the following major contributions to this important area of research: (i) We use the predicted future motion in the static image (Walker et al., 2015) as a means of compensating for the missing temporal information, while using the saliency map to represent the the spatial information in the form of location and shape of what is predicted as significant. (ii) We cast action classification in static images as a domain adaptation problem by transfer learning. We first map the input static image to a new domain that we refer to as the Predicted Optical Flow-Saliency Map domain (POF-SM), and then fine-tune the layers of a deep CNN model trained on classifying the ImageNet dataset to perform action classification in the POF-SM domain. (iii) We tested our method on the popular Willow dataset. But unlike existing methods, we also tested on a more realistic and challenging dataset of over 2M still images that we collected and labeled by taking random frames from the UCF-101 video dataset. We call our dataset the UCF Still Image dataset or UCFSI-101 in short. Our results outperform the state of the art.
Estimating action quality, the process of assigning a "score" to the execution of an action, is crucial in areas such as sports and health care. Unlike action recognition, which has millions of examples to learn from, the action quality datasets that are currently available are small -- typically comprised of only a few hundred samples. This work presents three frameworks for evaluating Olympic sports which utilize spatiotemporal features learned using 3D convolutional neural networks (C3D) and perform score regression with i) SVR, ii) LSTM, and iii) LSTM followed by SVR. An efficient training mechanism for the limited data scenarios is presented for clip-based training with LSTM. The proposed systems show significant improvement over existing quality assessment approaches on the task of predicting scores of Olympic events {diving, vault, figure skating}. While the SVR-based frameworks yield better results, LSTM-based frameworks are more natural for describing an action and can be used for improvement feedback.
Infrared (IR) imaging has the potential to enable more robust action recognition systems compared to visible spectrum cameras due to lower sensitivity to lighting conditions and appearance variability. While the action recognition task on videos collected from visible spectrum imaging has received much attention, action recognition in IR videos is significantly less explored. Our objective is to exploit imaging data in this modality for the action recognition task. In this work, we propose a novel two-stream 3D convolutional neural network (CNN) architecture by introducing the discriminative code layer and the corresponding discriminative code loss function. The proposed network processes IR image and the IR-based optical flow field sequences. We pretrain the 3D CNN model on the visible spectrum Sports-1M action dataset and finetune it on the Infrared Action Recognition (InfAR) dataset. To our best knowledge, this is the first application of the 3D CNN to action recognition in the IR domain. We conduct an elaborate analysis of different fusion schemes (weighted average, single and double-layer neural nets) applied to different 3D CNN outputs. Experimental results demonstrate that our approach can achieve state-of-the-art average precision (AP) performances on the InfAR dataset: (1) the proposed two-stream 3D CNN achieves the best reported 77.5% AP, and (2) our 3D CNN model applied to the optical flow fields achieves the best reported single stream 75.42% AP.
Recently, the soft attention mechanism, which was originally proposed in language processing, has been applied in computer vision tasks like image captioning. This paper presents improvements to the soft attention model by combining a convolutional LSTM with a hierarchical system architecture to recognize action categories in videos. We call this model the Convolutional Hierarchical Attention Model (CHAM). The model applies a convolutional operation inside the LSTM cell and an attention map generation process to recognize actions. The hierarchical architecture of this model is able to explicitly reason on multi-granularities of action categories. The proposed architecture achieved improved results on three publicly available datasets: the UCF sports dataset, the Olympic sports dataset and the HMDB51 dataset.
Current action recognition methods heavily rely on trimmed videos for model training. However, it is expensive and time-consuming to acquire a large-scale trimmed video dataset. This paper presents a new weakly supervised architecture, called UntrimmedNet, which is able to directly learn action recognition models from untrimmed videos without the requirement of temporal annotations of action instances. Our UntrimmedNet couples two important components, the classification module and the selection module, to learn the action models and reason about the temporal duration of action instances, respectively. These two components are implemented with feed-forward networks, and UntrimmedNet is therefore an end-to-end trainable architecture. We exploit the learned models for action recognition (WSR) and detection (WSD) on the untrimmed video datasets of THUMOS14 and ActivityNet. Although our UntrimmedNet only employs weak supervision, our method achieves performance superior or comparable to that of those strongly supervised approaches on these two datasets.
Shot boundary detection (SBD) is an important component of many video analysis tasks, such as action recognition, video indexing, summarization and editing. Previous work typically used a combination of low-level features like color histograms, in conjunction with simple models such as SVMs. Instead, we propose to learn shot detection end-to-end, from pixels to final shot boundaries. For training such a model, we rely on our insight that all shot boundaries are generated. Thus, we create a dataset with one million frames and automatically generated transitions such as cuts, dissolves and fades. In order to efficiently analyze hours of videos, we propose a Convolutional Neural Network (CNN) which is fully convolutional in time, thus allowing to use a large temporal context without the need to repeatedly processing frames. With this architecture our method obtains state-of-the-art results while running at an unprecedented speed of more than 120x real-time.
Representations that can compactly and effectively capture temporal evolution of semantic content are important to machine learning algorithms that operate on multi-variate time-series data. We investigate such representations motivated by the task of human action recognition. Here each data instance is encoded by a multivariate feature (such as via a deep CNN) where action dynamics are characterized by their variations in time. As these features are often non-linear, we propose a novel pooling method, kernelized rank pooling, that represents a given sequence compactly as the pre-image of the parameters of a hyperplane in an RKHS, projections of data onto which captures their temporal order. We develop this idea further and show that such a pooling scheme can be cast as an order-constrained kernelized PCA objective; we then propose to use the parameters of a kernelized low-rank feature subspace as the representation of the sequences. We cast our formulation as an optimization problem on generalized Grassmann manifolds and then solve it efficiently using Riemannian optimization techniques. We present experiments on several action recognition datasets using diverse feature modalities and demonstrate state-of-the-art results.
In many sports, it is useful to analyse video of an athlete in competition for training purposes. In swimming, stroke rate is a common metric used by coaches; requiring a laborious labelling of each individual stroke. We show that using a Convolutional Neural Network (CNN) we can automatically detect discrete events in continuous video (in this case, swimming strokes). We create a CNN that learns a mapping from a window of frames to a point on a smooth 1D target signal, with peaks denoting the location of a stroke, evaluated as a sliding window. To our knowledge this process of training and utilizing a CNN has not been investigated before; either in sports or fundamental computer vision research. Most research has been focused on action recognition and using it to classify many clips in continuous video for action localisation. In this paper we demonstrate our process works well on the task of detecting swimming strokes in the wild. However, without modifying the model architecture or training method, the process is also shown to work equally well on detecting tennis strokes, implying that this is a general process. The outputs of our system are surprisingly smooth signals that predict an arbitrary event at least as accurately as humans (manually evaluated from a sample of negative results). A number of different architectures are evaluated, pertaining to slightly different problem formulations and signal targets.
General human action recognition requires understanding of various visual cues. In this paper, we propose a network architecture that computes and integrates the most important visual cues for action recognition: pose, motion, and the raw images. For the integration, we introduce a Markov chain model which adds cues successively. The resulting approach is efficient and applicable to action classification as well as to spatial and temporal action localization. The two contributions clearly improve the performance over respective baselines. The overall approach achieves state-of-the-art action classification performance on HMDB51, J-HMDB and NTU RGB+D datasets. Moreover, it yields state-of-the-art spatio-temporal action localization results on UCF101 and J-HMDB.
Typical human actions last several seconds and exhibit characteristic spatio-temporal structure. Recent methods attempt to capture this structure and learn action representations with convolutional neural networks. Such representations, however, are typically learned at the level of a few video frames failing to model actions at their full temporal extent. In this work we learn video representations using neural networks with long-term temporal convolutions (LTC). We demonstrate that LTC-CNN models with increased temporal extents improve the accuracy of action recognition. We also study the impact of different low-level representations, such as raw values of video pixels and optical flow vector fields and demonstrate the importance of high-quality optical flow estimation for learning accurate action models. We report state-of-the-art results on two challenging benchmarks for human action recognition UCF101 (92.7%) and HMDB51 (67.2%).
Recently, Long Short-Term Memory (LSTM) has become a popular choice to model individual dynamics for single-person action recognition due to its ability of modeling the temporal information in various ranges of dynamic contexts. However, existing RNN models only focus on capturing the temporal dynamics of the person-person interactions by naively combining the activity dynamics of individuals or modeling them as a whole. This neglects the inter-related dynamics of how person-person interactions change over time. To this end, we propose a novel Concurrence-Aware Long Short-Term Sub-Memories (Co-LSTSM) to model the long-term inter-related dynamics between two interacting people on the bounding boxes covering people. Specifically, for each frame, two sub-memory units store individual motion information, while a concurrent LSTM unit selectively integrates and stores inter-related motion information between interacting people from these two sub-memory units via a new co-memory cell. Experimental results on the BIT and UT datasets show the superiority of Co-LSTSM compared with the state-of-the-art methods.
This paper presents a new method for 3D action recognition with skeleton sequences (i.e., 3D trajectories of human skeleton joints). The proposed method first transforms each skeleton sequence into three clips each consisting of several frames for spatial temporal feature learning using deep neural networks. Each clip is generated from one channel of the cylindrical coordinates of the skeleton sequence. Each frame of the generated clips represents the temporal information of the entire skeleton sequence, and incorporates one particular spatial relationship between the joints. The entire clips include multiple frames with different spatial relationships, which provide useful spatial structural information of the human skeleton. We propose to use deep convolutional neural networks to learn long-term temporal information of the skeleton sequence from the frames of the generated clips, and then use a Multi-Task Learning Network (MTLN) to jointly process all frames of the generated clips in parallel to incorporate spatial structural information for action recognition. Experimental results clearly show the effectiveness of the proposed new representation and feature learning method for 3D action recognition.
Zero-shot learning for visual recognition, e.g., object and action recognition, has recently attracted a lot of attention. However, it still remains challenging in bridging the semantic gap between visual features and their underlying semantics and transferring knowledge to semantic categories unseen during learning. Unlike most of the existing zero-shot visual recognition methods, we propose a stagewise bidirectional latent embedding framework to two subsequent learning stages for zero-shot visual recognition. In the bottom-up stage, a latent embedding space is first created by exploring the topological and labeling information underlying training data of known classes via a proper supervised subspace learning algorithm and the latent embedding of training data are used to form landmarks that guide embedding semantics underlying unseen classes into this learned latent space. In the top-down stage, semantic representations of unseen-class labels in a given label vocabulary are then embedded to the same latent space to preserve the semantic relatedness between all different classes via our proposed semi-supervised Sammon mapping with the guidance of landmarks. Thus, the resultant latent embedding space allows for predicting the label of a test instance with a simple nearest-neighbor rule. To evaluate the effectiveness of the proposed framework, we have conducted extensive experiments on four benchmark datasets in object and action recognition, i.e., AwA, CUB-200-2011, UCF101 and HMDB51. The experimental results under comparative studies demonstrate that our proposed approach yields the state-of-the-art performance under inductive and transductive settings.
It remains a challenge to efficiently extract spatialtemporal information from skeleton sequences for 3D human action recognition. Although most recent action recognition methods are based on Recurrent Neural Networks which present outstanding performance, one of the shortcomings of these methods is the tendency to overemphasize the temporal information. Since 3D convolutional neural network(3D CNN) is a powerful tool to simultaneously learn features from both spatial and temporal dimensions through capturing the correlations between three dimensional signals, this paper proposes a novel two-stream model using 3D CNN. To our best knowledge, this is the first application of 3D CNN in skeleton-based action recognition. Our method consists of three stages. First, skeleton joints are mapped into a 3D coordinate space and then encoding the spatial and temporal information, respectively. Second, 3D CNN models are seperately adopted to extract deep features from two streams. Third, to enhance the ability of deep features to capture global relationships, we extend every stream into multitemporal version. Extensive experiments on the SmartHome dataset and the large-scale NTU RGB-D dataset demonstrate that our method outperforms most of RNN-based methods, which verify the complementary property between spatial and temporal information and the robustness to noise.
Understanding the simultaneously very diverse and intricately fine-grained set of possible human actions is a critical open problem in computer vision. Manually labeling training videos is feasible for some action classes but doesn't scale to the full long-tailed distribution of actions. A promising way to address this is to leverage noisy data from web queries to learn new actions, using semi-supervised or "webly-supervised" approaches. However, these methods typically do not learn domain-specific knowledge, or rely on iterative hand-tuned data labeling policies. In this work, we instead propose a reinforcement learning-based formulation for selecting the right examples for training a classifier from noisy web search results. Our method uses Q-learning to learn a data labeling policy on a small labeled training dataset, and then uses this to automatically label noisy web data for new visual concepts. Experiments on the challenging Sports-1M action recognition benchmark as well as on additional fine-grained and newly emerging action classes demonstrate that our method is able to learn good labeling policies for noisy data and use this to learn accurate visual concept classifiers.
Every moment counts in action recognition. A comprehensive understanding of human activity in video requires labeling every frame according to the actions occurring, placing multiple labels densely over a video sequence. To study this problem we extend the existing THUMOS dataset and introduce MultiTHUMOS, a new dataset of dense labels over unconstrained internet videos. Modeling multiple, dense labels benefits from temporal relations within and across classes. We define a novel variant of long short-term memory (LSTM) deep networks for modeling these temporal relations via multiple input and output connections. We show that this model improves action labeling accuracy and further enables deeper understanding tasks ranging from structured retrieval to action prediction.
Automatic recognition of facial gestures is becoming increasingly important as real world AI agents become a reality. In this paper, we present an automated system that recognizes facial gestures by capturing local changes and encoding the motion into a histogram of frequencies. We evaluate the proposed method by demonstrating its effectiveness on spontaneous face action benchmarks: the FEEDTUM dataset, the Pain dataset and the HMDB51 dataset. The results show that, compared to known methods, the new encoding methods significantly improve the recognition accuracy and the robustness of analysis for a variety of applications.
State-of-the-art temporal action detectors inefficiently search the entire video for specific actions. Despite the encouraging progress these methods achieve, it is crucial to design automated approaches that only explore parts of the video which are the most relevant to the actions being searched for. To address this need, we propose the new problem of action spotting in video, which we define as finding a specific action in a video while observing a small portion of that video. Inspired by the observation that humans are extremely efficient and accurate in spotting and finding action instances in video, we propose Action Search, a novel Recurrent Neural Network approach that mimics the way humans spot actions. Moreover, to address the absence of data recording the behavior of human annotators, we put forward the Human Searches dataset, which compiles the search sequences employed by human annotators spotting actions in the AVA and THUMOS14 datasets. We consider temporal action localization as an application of the action spotting problem. Experiments on the THUMOS14 dataset reveal that our model is not only able to explore the video efficiently (observing on average 17.3% of the video) but it also accurately finds human activities with 30.8% mAP.
In this paper the problem of complex event detection in the continuous domain (i.e. events with unknown starting and ending locations) is addressed. Existing event detection methods are limited to features that are extracted from the local spatial or spatio-temporal patches from the videos. However, this makes the model vulnerable to the events with similar concepts e.g. "Open drawer" and "Open cupboard". In this work, in order to address the aforementioned limitations we present a novel model based on the combination of semantic and temporal features extracted from video frames. We train a max-margin classifier on top of the extracted features in an adaptive framework that is able to detect the events with unknown starting and ending locations. Our model is based on the Bidirectional Region Neural Network and large margin Structural Output SVM. The generality of our model allows it to be simply applied to different labeled and unlabeled datasets. We finally test our algorithm on three challenging datasets, "UCF 101-Action Recognition", "MPII Cooking Activities" and "Hollywood", and we report state-of-the-art performance.
This paper presents an image classification based approach for skeleton-based video action recognition problem. Firstly, A dataset independent translation-scale invariant image mapping method is proposed, which transformes the skeleton videos to colour images, named skeleton-images. Secondly, A multi-scale deep convolutional neural network (CNN) architecture is proposed which could be built and fine-tuned on the powerful pre-trained CNNs, e.g., AlexNet, VGGNet, ResNet etal.. Even though the skeleton-images are very different from natural images, the fine-tune strategy still works well. At last, we prove that our method could also work well on 2D skeleton video data. We achieve the state-of-the-art results on the popular benchmard datasets e.g. NTU RGB+D, UTD-MHAD, MSRC-12, and G3D. Especially on the largest and challenge NTU RGB+D, UTD-MHAD, and MSRC-12 dataset, our method outperforms other methods by a large margion, which proves the efficacy of the proposed method.