This technical report presents an overview of our solution used in the submission to ActivityNet Challenge 2019 Task 1 (\textbf{temporal action proposal generation}) and Task 2 (\textbf{temporal action localization/detection}). Temporal action proposal indicates the temporal intervals containing the actions and plays an important role in temporal action localization. Top-down and bottom-up methods are the two main categories used for proposal generation in the existing literature. In this paper, we devise a novel Multi-Granularity Fusion Network (MGFN) to combine the proposals generated from different frameworks for complementary filtering and confidence re-ranking. Specifically, we consider the diversity comprehensively from multiple perspectives, e.g. the characteristic aspect, the data aspect, the model aspect and the result aspect. Our MGFN achieves the state-of-the-art performance on the temporal action proposal task with 69.85 AUC score and the temporal action localization task with 38.90 mAP on the challenge testing set.
Human actions in videos are 3D signals. However, there are a few methods available for multiple human action recognition. For long videos, it's difficult to search within a video for a specific action and/or person. For that, this paper proposes a new technic for multiple human action recognition and summarization for surveillance videos. The proposed approach proposes a new representation of the data by extracting the sequence of each person from the scene. This is followed by an analysis of each sequence to detect and recognize the corresponding actions using 3D convolutional neural networks (3DCNNs). Action-based video summarization is performed by saving each person's action at each time of the video. Results of this work revealed that the proposed method provides accurate multi human action recognition that easily used for summarization of any action. Further, for other videos that can be collected from the internet, which are complex and not built for surveillance applications, the proposed model was evaluated on some datasets like UCF101 and YouTube without any preprocessing. For this category of videos, the summarization is performed on the video sequences by summarizing the actions in each subsequence. The results obtained demonstrate its efficiency compared to state-of-the-art methods.
This technical report present an overview of our system proposed for the spatio-temporal action localization(SAL) task in ActivityNet Challenge 2019. Unlike previous two-streams-based works, we focus on exploring the end-to-end trainable architecture using only RGB sequential images. To this end, we employ a previously proposed simple yet effective two-branches network called SlowFast Networks which is capable of capturing both short- and long-term spatiotemporal features. Moreover, to handle the severe class imbalance and overfitting problems, we propose a correlation-preserving data augmentation method and a random label subsampling method which have been proven to be able to reduce overfitting and improve the performance.
Temporal action proposal generation is an challenging and promising task which aims to locate temporal regions in real-world videos where action or event may occur. Current bottom-up proposal generation methods can generate proposals with precise boundary, but cannot efficiently generate adequately reliable confidence scores for retrieving proposals. To address these difficulties, we introduce the Boundary-Matching (BM) mechanism to evaluate confidence scores of densely distributed proposals, which denote a proposal as a matching pair of starting and ending boundaries and combine all densely distributed BM pairs into the BM confidence map. Based on BM mechanism, we propose an effective, efficient and end-to-end proposal generation method, named Boundary-Matching Network (BMN), which generates proposals with precise temporal boundaries as well as reliable confidence scores simultaneously. The two-branches of BMN are jointly trained in an unified framework. We conduct experiments on two challenging datasets: THUMOS-14 and ActivityNet-1.3, where BMN shows significant performance improvement with remarkable efficiency and generalizability. Further, combining with existing action classifier, BMN can achieve state-of-the-art temporal action detection performance.
In this paper we propose a novel Temporal Attentive Relation Network (TARN) for the problems of few-shot and zero-shot action recognition. At the heart of our network is a meta-learning approach that learns to compare representations of variable temporal length, that is, either two videos of different length (in the case of few-shot action recognition) or a video and a semantic representation such as word vector (in the case of zero-shot action recognition). By contrast to other works in few-shot and zero-shot action recognition, we a) utilise attention mechanisms so as to perform temporal alignment, and b) learn a deep-distance measure on the aligned representations at video segment level. We adopt an episode-based training scheme and train our network in an end-to-end manner. The proposed method does not require any fine-tuning in the target domain or maintaining additional representations as is the case of memory networks. Experimental results show that the proposed architecture outperforms the state of the art in few-shot action recognition, and achieves competitive results in zero-shot action recognition.
Slow Feature Analysis (SFA) extracts slowly varying features from a quickly varying input signal. It has been successfully applied to modeling the visual receptive fields of the cortical neurons. Sufficient experimental results in neuroscience suggest that the temporal slowness principle is a general learning principle in visual perception. In this paper, we introduce the SFA framework to the problem of human action recognition by incorporating the discriminative information with SFA learning and considering the spatial relationship of body parts. In particular, we consider four kinds of SFA learning strategies, including the original unsupervised SFA (U-SFA), the supervised SFA (S-SFA), the discriminative SFA (D-SFA), and the spatial discriminative SFA (SD-SFA), to extract slow feature functions from a large amount of training cuboids which are obtained by random sampling in motion boundaries. Afterward, to represent action sequences, the squared first order temporal derivatives are accumulated over all transformed cuboids into one feature vector, which is termed the Accumulated Squared Derivative (ASD) feature. The ASD feature encodes the statistical distribution of slow features in an action sequence. Finally, a linear support vector machine (SVM) is trained to classify actions represented by ASD features. We conduct extensive experiments, including two sets of control experiments, two sets of large scale experiments on the KTH and Weizmann databases, and two sets of experiments on the CASIA and UT-interaction databases, to demonstrate the effectiveness of SFA for human action recognition.
Understanding the content of videos is one of the core techniques for developing various helpful applications in the real world, such as recognizing various human actions for surveillance systems or customer behavior analysis in an autonomous shop. However, understanding the content or story of the video still remains a challenging problem due to its sheer amount of data and temporal structure. In this paper, we propose a multi-channel neural network structure that adopts a two-stream network structure, which has been shown high performance in human action recognition field, and use it as a spatiotemporal video feature extractor for solving video question and answering task. We also adopt a squeeze-and-excitation structure to two-stream network structure for achieving a channel-wise attended spatiotemporal feature. For jointly modeling the spatiotemporal features from video and the textual features from the question, we design a context matching module with a level adjusting layer to remove the gap of information between visual and textual features by applying attention mechanism on joint modeling. Finally, we adopt a scoring mechanism and smoothed ranking loss objective function for selecting the correct answer from answer candidates. We evaluate our model with TVQA dataset, and our approach shows the improved result in textual only setting, but the result with visual feature shows the limitation and possibility of our approach.
Deep neural networks have achieved great success for video analysis and understanding. However, designing a high-performance neural architecture requires substantial efforts and expertise. In this paper, we make the first attempt to let algorithm automatically design neural networks for video action recognition tasks. Specifically, a spatio-temporal network is developed in a differentiable space modeled by a directed acyclic graph, thus a gradient-based strategy can be performed to search an optimal architecture. Nonetheless, it is computationally expensive, since the computational burden to evaluate each architecture candidate is still heavy. To alleviate this issue, we, for the video input, introduce a temporal segment approach to reduce the computational cost without losing global video information. For the architecture, we explore in an efficient search space by introducing pseudo 3D operators. Experiments show that, our architecture outperforms popular neural architectures, under the training from scratch protocol, on the challenging UCF101 dataset, surprisingly, with only around one percentage of parameters of its manual-design counterparts.
We present a new deep learning approach for real-time 3D human action recognition from skeletal data and apply it to develop a vision-based intelligent surveillance system. Given a skeleton sequence, we propose to encode skeleton poses and their motions into a single RGB image. An Adaptive Histogram Equalization (AHE) algorithm is then applied on the color images to enhance their local patterns and generate more discriminative features. For learning and classification tasks, we design Deep Neural Networks based on the Densely Connected Convolutional Architecture (DenseNet) to extract features from enhanced-color images and classify them into classes. Experimental results on two challenging datasets show that the proposed method reaches state-of-the-art accuracy, whilst requiring low computational time for training and inference. This paper also introduces CEMEST, a new RGB-D dataset depicting passenger behaviors in public transport. It consists of 203 untrimmed real-world surveillance videos of realistic normal and anomalous events. We achieve promising results on real conditions of this dataset with the support of data augmentation and transfer learning techniques. This enables the construction of real-world applications based on deep learning for enhancing monitoring and security in public transport.
Video-based human action recognition is currently one of the most active research areas in computer vision. Various research studies indicate that the performance of action recognition is highly dependent on the type of features being extracted and how the actions are represented. Since the release of the Kinect camera, a large number of Kinect-based human action recognition techniques have been proposed in the literature. However, there still does not exist a thorough comparison of these Kinect-based techniques under the grouping of feature types, such as handcrafted versus deep learning features and depth-based versus skeleton-based features. In this paper, we analyze and compare ten recent Kinect-based algorithms for both cross-subject action recognition and cross-view action recognition using six benchmark datasets. In addition, we have implemented and improved some of these techniques and included their variants in the comparison. Our experiments show that the majority of methods perform better on cross-subject action recognition than cross-view action recognition, that skeleton-based features are more robust for cross-view recognition than depth-based features, and that deep learning features are suitable for large datasets.
Human visual system can selectively attend to parts of a scene for quick perception, a biological mechanism known as Human attention. Inspired by this, recent deep learning models encode attention mechanisms to focus on the most task-relevant parts of the input signal for further processing, which is called Machine/Neural/Artificial attention. Understanding the relation between human and machine attention is important for interpreting and designing neural networks. Many works claim that the attention mechanism offers an extra dimension of interpretability by explaining where the neural networks look. However, recent studies demonstrate that artificial attention maps do not always coincide with common intuition. In view of these conflicting evidence, here we make a systematic study on using artificial attention and human attention in neural network design. With three example computer vision tasks, diverse representative backbones, and famous architectures, corresponding real human gaze data, and systematically conducted large-scale quantitative studies, we quantify the consistency between artificial attention and human visual attention and offer novel insights into existing artificial attention mechanisms by giving preliminary answers to several key questions related to human and artificial attention mechanisms. Overall results demonstrate that human attention can benchmark the meaningful `ground-truth' in attention-driven tasks, where the more the artificial attention is close to human attention, the better the performance; for higher-level vision tasks, it is case-by-case. It would be advisable for attention-driven tasks to explicitly force a better alignment between artificial and human attention to boost the performance; such alignment would also improve the network explainability for higher-level computer vision tasks.
This notebook paper presents an overview and comparative analysis of our systems designed for the following three tasks in ActivityNet Challenge 2019: trimmed action recognition, dense-captioning events in videos, and spatio-temporal action localization.
Convolutional Neural Networks (CNN) have been regarded as a powerful class of models for visual recognition problems. Nevertheless, the convolutional filters in these networks are local operations while ignoring the large-range dependency. Such drawback becomes even worse particularly for video recognition, since video is an information-intensive media with complex temporal variations. In this paper, we present a novel framework to boost the spatio-temporal representation learning by Local and Global Diffusion (LGD). Specifically, we construct a novel neural network architecture that learns the local and global representations in parallel. The architecture is composed of LGD blocks, where each block updates local and global features by modeling the diffusions between these two representations. Diffusions effectively interact two aspects of information, i.e., localized and holistic, for more powerful way of representation learning. Furthermore, a kernelized classifier is introduced to combine the representations from two aspects for video recognition. Our LGD networks achieve clear improvements on the large-scale Kinetics-400 and Kinetics-600 video classification datasets against the best competitors by 3.5% and 0.7%. We further examine the generalization of both the global and local representations produced by our pre-trained LGD networks on four different benchmarks for video action recognition and spatio-temporal action detection tasks. Superior performances over several state-of-the-art techniques on these benchmarks are reported. Code is available at: https://github.com/ZhaofanQiu/local-and-global-diffusion-networks.
Action recognition is so far mainly focusing on the problem of classification of hand selected preclipped actions and reaching impressive results in this field. But with the performance even ceiling on current datasets, it also appears that the next steps in the field will have to go beyond this fully supervised classification. One way to overcome those problems is to move towards less restricted scenarios. In this context we present a large-scale real-world dataset designed to evaluate learning techniques for human action recognition beyond hand-crafted datasets. To this end we put the process of collecting data on its feet again and start with the annotation of a test set of 250 cooking videos. The training data is then gathered by searching for the respective annotated classes within the subtitles of freely available videos. The uniqueness of the dataset is attributed to the fact that the whole process of collecting the data and training does not involve any human intervention. To address the problem of semantic inconsistencies that arise with this kind of training data, we further propose a semantical hierarchical structure for the mined classes.
Inspired by the observation that humans are able to process videos efficiently by only paying attention where and when it is needed, we propose an interpretable and easy plug-in spatial-temporal attention mechanism for video action recognition. For spatial attention, we learn a saliency mask to allow the model to focus on the most salient parts of the feature maps. For temporal attention, we employ a convolutional LSTM based attention mechanism to identify the most relevant frames from an input video. Further, we propose a set of regularizers to ensure that our attention mechanism attends to coherent regions in space and time. Our model not only improves video action recognition accuracy, but also localizes discriminative regions both spatially and temporally, despite being trained in a weakly-supervised manner with only classification labels (no bounding box labels or time frame temporal labels). We evaluate our approach on several public video action recognition datasets with ablation studies. Furthermore, we quantitatively and qualitatively evaluate our model's ability to localize discriminative regions spatially and critical frames temporally. Experimental results demonstrate the efficacy of our approach, showing superior or comparable accuracy with the state-of-the-art methods while increasing model interpretability.
Consider end-to-end training of a multi-modal vs. a single-modal network on a task with multiple input modalities: the multi-modal network receives more information, so it should match or outperform its single-modal counterpart. In our experiments, however, we observe the opposite: the best single-modal network always outperforms the multi-modal network. This observation is consistent across different combinations of modalities and on different tasks and benchmarks. This paper identifies two main causes for this performance drop: first, multi-modal networks are often prone to overfitting due to increased capacity. Second, different modalities overfit and generalize at different rates, so training them jointly with a single optimization strategy is sub-optimal. We address these two problems with a technique we call Gradient Blending, which computes an optimal blend of modalities based on their overfitting behavior. We demonstrate that Gradient Blending outperforms widely-used baselines for avoiding overfitting and achieves state-of-the-art accuracy on various tasks including human action recognition, ego-centric action recognition, and acoustic event detection.
Recurrent neural networks (RNNs) are omnipresent in sequence modeling tasks. Practical models usually consist of several layers of hundreds or thousands of neurons which are fully connected. This places a heavy computational and memory burden on hardware, restricting adoption in practical low-cost and low-power devices. Compared to fully convolutional models, the costly sequential operation of RNNs severely hinders performance on parallel hardware. This paper challenges the convention of full connectivity in RNNs. We study structurally sparse RNNs, showing that they are well suited for acceleration on parallel hardware, with a greatly reduced cost of the recurrent operations as well as orders of magnitude less recurrent weights. Extensive experiments on challenging tasks ranging from language modeling and speech recognition to video action recognition reveal that structurally sparse RNNs achieve competitive performance as compared to fully-connected networks. This allows for using large sparse RNNs for a wide range of real-world tasks that previously were too costly with fully connected networks.
Most action recognition methods base on a) a late aggregation of frame level CNN features using average pooling, max pooling, or RNN, among others, or b) spatio-temporal aggregation via 3D convolutions. The first assume independence among frame features up to a certain level of abstraction and then perform higher-level aggregation, while the second extracts spatio-temporal features from grouped frames as early fusion. In this paper we explore the space in between these two, by letting adjacent feature branches interact as they develop into the higher level representation. The interaction happens between feature differencing and averaging at each level of the hierarchy, and it has convolutional structure that learns to select the appropriate mode locally in contrast to previous works that impose one of the modes globally (e.g. feature differencing) as a design choice. We further constrain this interaction to be conservative, e.g. a local feature subtraction in one branch is compensated by the addition on another, such that the total feature flow is preserved. We evaluate the performance of our proposal on a number of existing models, i.e. TSN, TRN and ECO, to show its flexibility and effectiveness in improving action recognition performance.
Spatio-temporal action localization consists of three levels of tasks: spatial localization, action classification, and temporal segmentation. In this work, we propose a new Progressive Cross-stream Cooperation (PCSC) framework to use both region proposals and features from one stream (i.e. Flow/RGB) to help another stream (i.e. RGB/Flow) to iteratively improve action localization results and generate better bounding boxes in an iterative fashion. Specifically, we first generate a larger set of region proposals by combining the latest region proposals from both streams, from which we can readily obtain a larger set of labelled training samples to help learn better action detection models. Second, we also propose a new message passing approach to pass information from one stream to another stream in order to learn better representations, which also leads to better action detection models. As a result, our iterative framework progressively improves action localization results at the frame level. To improve action localization results at the video level, we additionally propose a new strategy to train class-specific actionness detectors for better temporal segmentation, which can be readily learnt by focusing on "confusing" samples from the same action class. Comprehensive experiments on two benchmark datasets UCF-101-24 and J-HMDB demonstrate the effectiveness of our newly proposed approaches for spatio-temporal action localization in realistic scenarios.
Temporal action localization has recently attracted significant interest in the Computer Vision community. However, despite the great progress, it is hard to identify which aspects of the proposed methods contribute most to the increase in localization performance. To address this issue, we conduct ablative experiments on feature extraction methods, fixed-size feature representation methods and training strategies, and report how each influences the overall performance. Based on our findings, we propose a two-stage detector that outperforms the state of the art in THUMOS14, achieving a mAP@tIoU=0.5 equal to 44.2%.