Temporal action localization is an important task of computer vision. Though a variety of methods have been proposed, it still remains an open question how to predict the temporal boundaries of action segments precisely. Most works use segment-level classifiers to select video segments pre-determined by action proposal or dense sliding windows. However, in order to achieve more precise action boundaries, a temporal localization system should make dense predictions at a fine granularity. A newly proposed work exploits Convolutional-Deconvolutional-Convolutional (CDC) filters to upsample the predictions of 3D ConvNets, making it possible to perform per-frame action predictions and achieving promising performance in terms of temporal action localization. However, CDC network loses temporal information partially due to the temporal downsampling operation. In this paper, we propose an elegant and powerful Temporal Preservation Convolutional (TPC) Network that equips 3D ConvNets with TPC filters. TPC network can fully preserve temporal resolution and downsample the spatial resolution simultaneously, enabling frame-level granularity action localization. TPC network can be trained in an end-to-end manner. Experiment results on public datasets show that TPC network achieves significant improvement on per-frame action prediction and competing results on segment-level temporal action localization.
For the two-stream style methods in action recognition, fusing the two streams' predictions is always by the weighted averaging scheme. This fusion method with fixed weights lacks of pertinence to different action videos and always needs trial and error on the validation set. In order to enhance the adaptability of two-stream ConvNets and improve its performance, an end-to-end trainable gated fusion method, namely gating ConvNet, for the two-stream ConvNets is proposed in this paper based on the MoE (Mixture of Experts) theory. The gating ConvNet takes the combination of feature maps from the same layer of the spatial and the temporal nets as input and adopts ReLU (Rectified Linear Unit) as the gating output activation function. To reduce the over-fitting of gating ConvNet caused by the redundancy of parameters, a new multi-task learning method is designed, which jointly learns the gating fusion weights for the two streams and learns the gating ConvNet for action classification. With our gated fusion method and multi-task learning approach, a high accuracy of 94.5% is achieved on the dataset UCF101.
Human action recognition refers to automatic recognizing human actions from a video clip. In reality, there often exist multiple human actions in a video stream. Such a video stream is often weakly-annotated with a set of relevant human action labels at a global level rather than assigning each label to a specific video episode corresponding to a single action, which leads to a multi-label learning problem. Furthermore, there are many meaningful human actions in reality but it would be extremely difficult to collect/annotate video clips regarding all of various human actions, which leads to a zero-shot learning scenario. To the best of our knowledge, there is no work that has addressed all the above issues together in human action recognition. In this paper, we formulate a real-world human action recognition task as a multi-label zero-shot learning problem and propose a framework to tackle this problem in a holistic way. Our framework holistically tackles the issue of unknown temporal boundaries between different actions for multi-label learning and exploits the side information regarding the semantic relationship between different human actions for knowledge transfer. Consequently, our framework leads to a joint latent ranking embedding for multi-label zero-shot human action recognition. A novel neural architecture of two component models and an alternate learning algorithm are proposed to carry out the joint latent ranking embedding learning. Thus, multi-label zero-shot recognition is done by measuring relatedness scores of action labels to a test video clip in the joint latent visual and semantic embedding spaces. We evaluate our framework with different settings, including a novel data split scheme designed especially for evaluating multi-label zero-shot learning, on two datasets: Breakfast and Charades. The experimental results demonstrate the effectiveness of our framework.
For effective human-robot interaction, it is important that a robotic assistant can forecast the next action a human will consider in a given task. Unfortunately, real-world tasks are often very long, complex, and repetitive; as a result forecasting is not trivial. In this paper, we propose a novel deep recurrent architecture that takes as input features from a two-stream Residual action recognition framework, and learns to estimate the progress of human activities from video sequences -- this surrogate progress estimation task implicitly learns a temporal task grammar with respect to which activities can be localized and forecasted. To learn the task grammar, we propose a stacked LSTM based multi-granularity progress estimation framework that uses a novel cumulative Euclidean loss as objective. To demonstrate the effectiveness of our proposed architecture, we showcase experiments on two challenging robotic assistive tasks, namely (i) assembling an Ikea table from its constituents, and (ii) changing the tires of a car. Our results demonstrate that learning task grammars offers highly discriminative cues improving the forecasting accuracy by more than 9% over the baseline two-stream forecasting model, while also outperforming other competitive schemes.
The recognition of human activities is one of the key problems in video understanding. Action recognition is challenging even for specific categories of videos, such as sports, that contain only a small set of actions. Interestingly, sports videos are accompanied by detailed commentaries available online, which could be used to perform action annotation in a weakly-supervised setting. For the specific case of Cricket videos, we address the challenge of temporal segmentation and annotation of ctions with semantic descriptions. Our solution consists of two stages. In the first stage, the video is segmented into "scenes", by utilizing the scene category information extracted from text-commentary. The second stage consists of classifying video-shots as well as the phrases in the textual description into various categories. The relevant phrases are then suitably mapped to the video-shots. The novel aspect of this work is the fine temporal scale at which semantic information is assigned to the video. As a result of our approach, we enable retrieval of specific actions that last only a few seconds, from several hours of video. This solution yields a large number of labeled exemplars, with no manual effort, that could be used by machine learning algorithms to learn complex actions.
Symmetric positive definite (SPD) matrices (e.g., covariances, graph Laplacians, etc.) are widely used to model the relationship of spatial or temporal domain. Nevertheless, SPD matrices are theoretically embedded on Riemannian manifolds. In this paper, we propose an end-to-end deep manifold-to-manifold transforming network (DMT-Net) which can make SPD matrices flow from one Riemannian manifold to another more discriminative one. To learn discriminative SPD features characterizing both spatial and temporal dependencies, we specifically develop three novel layers on manifolds: (i) the local SPD convolutional layer, (ii) the non-linear SPD activation layer, and (iii) the Riemannian-preserved recursive layer. The SPD property is preserved through all layers without any requirement of singular value decomposition (SVD), which is often used in the existing methods with expensive computation cost. Furthermore, a diagonalizing SPD layer is designed to efficiently calculate the final metric for the classification task. To evaluate our proposed method, we conduct extensive experiments on the task of action recognition, where input signals are popularly modeled as SPD matrices. The experimental results demonstrate that our DMT-Net is much more competitive over state-of-the-art.
3D action recognition was shown to benefit from a covariance representation of the input data (joint 3D positions). A kernel machine feed with such feature is an effective paradigm for 3D action recognition, yielding state-of-the-art results. Yet, the whole framework is affected by the well-known scalability issue. In fact, in general, the kernel function has to be evaluated for all pairs of instances inducing a Gram matrix whose complexity is quadratic in the number of samples. In this work we reduce such complexity to be linear by proposing a novel and explicit feature map to approximate the kernel function. This allows to train a linear classifier with an explicit feature encoding, which implicitly implements a Log-Euclidean machine in a scalable fashion. Not only we prove that the proposed approximation is unbiased, but also we work out an explicit strong bound for its variance, attesting a theoretical superiority of our approach with respect to existing ones. Experimentally, we verify that our representation provides a compact encoding and outperforms other approximation schemes on a number of publicly available benchmark datasets for 3D action recognition.
Action recognition, motion classification, gait analysis and synthesis are fundamental problems in a number of fields such as computer graphics, bio-mechanics and human computer interaction that generate a large body of research. This type of data is complex because it is inherently multidimensional and has multiple modalities such as video, motion capture data, accelerometer data, etc. While some of this data, such as monocular video are easy to acquire, others are much more difficult and expensive such as motion capture data or multi-view video. This creates a large barrier of entry in the research community for data driven research. We have embarked on creating a new large repository of motion and action data (CAMREP) consisting of several motion and action databases. What makes this database unique is that we use a variety of modalities, enabling multi-modal analysis. Presently, the size of datasets varies with some having a large number of subjects while others having smaller numbers. We have also acquired long capture sequences in a number of cases, making some datasets rather large.
When we say a person is texting, can you tell the person is walking or sitting? Emphatically, no. In order to solve this incomplete representation problem, this paper presents a sub-action descriptor for detailed action detection. The sub-action descriptor consists of three levels: the posture, the locomotion, and the gesture level. The three levels give three sub-action categories for one action to address the representation problem. The proposed action detection model simultaneously localizes and recognizes the actions of multiple individuals in video surveillance using appearance-based temporal features with multi-CNN. The proposed approach achieved a mean average precision (mAP) of 76.6% at the frame-based and 83.5% at the video-based measurement on the new large-scale ICVL video surveillance dataset that the authors introduce and make available to the community with this paper. Extensive experiments on the benchmark KTH dataset demonstrate that the proposed approach achieved better performance, which in turn boosts the action recognition performance over the state-of-the-art. The action detection model can run at around 25 fps on the ICVL and more than 80 fps on the KTH dataset, which is suitable for real-time surveillance applications.
A fundamental challenge in machine learning today is to build a model that can learn from few examples. Here, we describe a reservoir based spiking neural model for learning to recognize actions with a limited number of labeled videos. First, we propose a novel encoding, inspired by how microsaccades influence visual perception, to extract spike information from raw video data while preserving the temporal correlation across different frames. Using this encoding, we show that the reservoir generalizes its rich dynamical activity toward signature action/movements enabling it to learn from few training examples. We evaluate our approach on the UCF-101 dataset. Our experiments demonstrate that our proposed reservoir achieves 81.3%/87% Top-1/Top-5 accuracy, respectively, on the 101-class data while requiring just 8 video examples per class for training. Our results establish a new benchmark for action recognition from limited video examples for spiking neural models while yielding competetive accuracy with respect to state-of-the-art non-spiking neural models.
Many computer vision applications involve modeling complex spatio-temporal patterns in high-dimensional motion data. Recently, restricted Boltzmann machines (RBMs) have been widely used to capture and represent spatial patterns in a single image or temporal patterns in several time slices. To model global dynamics and local spatial interactions, we propose to theoretically extend the conventional RBMs by introducing another term in the energy function to explicitly model the local spatial interactions in the input data. A learning method is then proposed to perform efficient learning for the proposed model. We further introduce a new method for multi-class classification that can effectively estimate the infeasible partition functions of different RBMs such that RBM is treated as a generative model for classification purpose. The improved RBM model is evaluated on two computer vision applications: facial expression recognition and human action recognition. Experimental results on benchmark databases demonstrate the effectiveness of the proposed algorithm.
Action recognition in surveillance video makes our life safer by detecting the criminal events or predicting violent emergencies. However, efficient action recognition is not free of difficulty. First, there are so many action classes in daily life that we cannot pre-define all possible action classes beforehand. Moreover, it is very hard to collect real-word videos for certain particular actions such as steal and street fight due to legal restrictions and privacy protection. These challenges make existing data-driven recognition methods insufficient to attain desired performance. Zero-shot learning is potential to be applied to solve these issues since it can perform classification without positive example. Nevertheless, current zero-shot learning algorithms have been studied under the unreasonable setting where seen classes are absent during the testing phase. Motivated by this, we study the task of action recognition in surveillance video under a more realistic \emph{generalized zero-shot setting}, where testing data contains both seen and unseen classes. To our best knowledge, this is the first work to study video action recognition under the generalized zero-shot setting. We firstly perform extensive empirical studies on several existing zero-shot leaning approaches under this new setting on a web-scale video data. Our experimental results demonstrate that, under the generalize setting, typical zero-shot learning methods are no longer effective for the dataset we applied. Then, we propose a method for action recognition by deploying generalized zero-shot learning, which transfers the knowledge of web video to detect the anomalous actions in surveillance videos. To verify the effectiveness of our proposed method, we further construct a new surveillance video dataset consisting of nine action classes related to the public safety situation.
This paper proposes a network architecture to perform variable length semantic video generation using captions. We adopt a new perspective towards video generation where we allow the captions to be combined with the long-term and short-term dependencies between video frames and thus generate a video in an incremental manner. Our experiments demonstrate our network architecture's ability to distinguish between objects, actions and interactions in a video and combine them to generate videos for unseen captions. The network also exhibits the capability to perform spatio-temporal style transfer when asked to generate videos for a sequence of captions. We also show that the network's ability to learn a latent representation allows it generate videos in an unsupervised manner and perform other tasks such as action recognition. (Accepted in International Conference in Computer Vision (ICCV) 2017)
Analyzing videos of human actions involves understanding the temporal relationships among video frames. State-of-the-art action recognition approaches rely on traditional optical flow estimation methods to pre-compute motion information for CNNs. Such a two-stage approach is computationally expensive, storage demanding, and not end-to-end trainable. In this paper, we present a novel CNN architecture that implicitly captures motion information between adjacent frames. We name our approach hidden two-stream CNNs because it only takes raw video frames as input and directly predicts action classes without explicitly computing optical flow. Our end-to-end approach is 10x faster than its two-stage baseline. Experimental results on four challenging action recognition datasets: UCF101, HMDB51, THUMOS14 and ActivityNet v1.2 show that our approach significantly outperforms the previous best real-time approaches.
Egocentric videos, which mainly record the activities carried out by the users of the wearable cameras, have drawn much research attentions in recent years. Due to its lengthy content, a large number of ego-related applications have been developed to abstract the captured videos. As the users are accustomed to interacting with the target objects using their own hands while their hands usually appear within their visual fields during the interaction, an egocentric hand detection step is involved in tasks like gesture recognition, action recognition and social interaction understanding. In this work, we propose a dynamic region growing approach for hand region detection in egocentric videos, by jointly considering hand-related motion and egocentric cues. We first determine seed regions that most likely belong to the hand, by analyzing the motion patterns across successive frames. The hand regions can then be located by extending from the seed regions, according to the scores computed for the adjacent superpixels. These scores are derived from four egocentric cues: contrast, location, position consistency and appearance continuity. We discuss how to apply the proposed method in real-life scenarios, where multiple hands irregularly appear and disappear from the videos. Experimental results on public datasets show that the proposed method achieves superior performance compared with the state-of-the-art methods, especially in complicated scenarios.
Image understanding using deep convolutional network has reached human-level performance, yet a closely related problem of video understanding especially, action recognition has not reached the requisite level of maturity. We combine multi-kernels based support-vector-machines (SVM) with a multi-stream deep convolutional neural network to achieve close to state-of-the-art performance on a 51-class activity recognition problem (HMDB-51 dataset); this specific dataset has proved to be particularly challenging for deep neural networks due to the heterogeneity in camera viewpoints, video quality, etc. The resulting architecture is named pillar networks as each (very) deep neural network acts as a pillar for the hierarchical classifiers. In addition, we illustrate that hand-crafted features such as improved dense trajectories (iDT) and Multi-skip Feature Stacking (MIFS), as additional pillars, can further supplement the performance.
The motion analysis of human skeletons is crucial for human action recognition, which is one of the most active topics in computer vision. In this paper, we propose a fully end-to-end action-attending graphic neural network (A$^2$GNN) for skeleton-based action recognition, in which each irregular skeleton is structured as an undirected attribute graph. To extract high-level semantic representation from skeletons, we perform the local spectral graph filtering on the constructed attribute graphs like the standard image convolution operation. Considering not all joints are informative for action analysis, we design an action-attending layer to detect those salient action units (AUs) by adaptively weighting skeletal joints. Herein the filtering responses are parameterized into a weighting function irrelevant to the order of input nodes. To further encode continuous motion variations, the deep features learnt from skeletal graphs are gathered along consecutive temporal slices and then fed into a recurrent gated network. Finally, the spectral graph filtering, action-attending and recurrent temporal encoding are integrated together to jointly train for the sake of robust action recognition as well as the intelligibility of human actions. To evaluate our A$^2$GNN, we conduct extensive experiments on four benchmark skeleton-based action datasets, including the large-scale challenging NTU RGB+D dataset. The experimental results demonstrate that our network achieves the state-of-the-art performances.
Symmetric positive definite (SPD) matrices used as feature descriptors in image recognition are usually high dimensional. Traditional manifold learning is only applicable for reducing the dimension of high-dimensional vector-form data. For high-dimensional SPD matrices, directly using manifold learning algorithms to reduce the dimension of matrix-form data is impossible. The SPD matrix must first be transformed into a long vector, and then the dimension of this vector must be reduced. However, this approach breaks the spatial structure of the SPD matrix space. To overcome this limitation, we propose a new dimension reduction algorithm on SPD matrix space to transform high-dimensional SPD matrices into low-dimensional SPD matrices. Our work is based on the fact that the set of all SPD matrices with the same size has a Lie group structure, and we aim to transform the manifold learning to the SPD matrix Lie group. We use the basic idea of the manifold learning algorithm called locality preserving projection (LPP) to construct the corresponding Laplacian matrix on the SPD matrix Lie group. Thus, we call our approach Lie-LPP to emphasize its Lie group character. We present a detailed algorithm analysis and show through experiments that Lie-LPP achieves effective results on human action recognition and human face recognition.
Action recognition is an important yet challenging task in computer vision. In this paper, we propose a novel deep-based framework for action recognition, which improves the recognition accuracy by: 1) deriving more precise features for representing actions, and 2) reducing the asynchrony between different information streams. We first introduce a coarse-to-fine network which extracts shared deep features at different action class granularities and progressively integrates them to obtain a more accurate feature representation for input actions. We further introduce an asynchronous fusion network. It fuses information from different streams by asynchronously integrating stream-wise features at different time points, hence better leveraging the complementary information in different streams. Experimental results on action recognition benchmarks demonstrate that our approach achieves the state-of-the-art performance.
Despite the growing discriminative capabilities of modern deep learning methods for recognition tasks, the inner workings of the state-of-art models still remain mostly black-boxes. In this paper, we propose a systematic interpretation of model parameters and hidden representations of Residual Temporal Convolutional Networks (Res-TCN) for action recognition in time-series data. We also propose a Feature Map Decoder as part of the interpretation analysis, which outputs a representation of model's hidden variables in the same domain as the input. Such analysis empowers us to expose model's characteristic learning patterns in an interpretable way. For example, through the diagnosis analysis, we discovered that our model has learned to achieve view-point invariance by implicitly learning to perform rotational normalization of the input to a more discriminative view. Based on the findings from the model interpretation analysis, we propose a targeted refinement technique, which can generalize to various other recognition models. The proposed work introduces a three-stage paradigm for model learning: training, interpretable diagnosis and targeted refinement. We validate our approach on skeleton based 3D human action recognition benchmark of NTU RGB+D. We show that the proposed workflow is an effective model learning strategy and the resulting Multi-stream Residual Temporal Convolutional Network (MS-Res-TCN) achieves the state-of-the-art performance on NTU RGB+D.