Zheng Shou

Zhijian Hou, Wanjun Zhong, Lei Ji et al.

This paper tackles an emerging and challenging problem of long video temporal grounding~(VTG) that localizes video moments related to a natural language (NL) query. Compared with short videos, long videos are also highly demanded but less explored, which brings new challenges in higher inference computation cost and weaker multi-modal alignment. To address these challenges, we propose CONE, an efficient COarse-to-fiNE alignment framework. CONE is a plug-and-play framework on top of existing VTG models to handle long videos through a sliding window mechanism. Specifically, CONE (1) introduces a query-guided window selection strategy to speed up inference, and (2) proposes a coarse-to-fine mechanism via a novel incorporation of contrastive learning to enhance multi-modal alignment for long videos. Extensive experiments on two large-scale long VTG benchmarks consistently show both substantial performance gains (e.g., from 3.13% to 6.87% on MAD) and state-of-the-art results. Analyses also reveal higher efficiency as the query-guided window selection mechanism accelerates inference time by 2x on Ego4D-NLQ and 15x on MAD while keeping SOTA results. Codes have been released at https://github.com/houzhijian/CONE.

Zhijian Hou, Wanjun Zhong, Lei Ji et al.

This technical report describes the CONE approach for Ego4D Natural Language Queries (NLQ) Challenge in ECCV 2022. We leverage our model CONE, an efficient window-centric COarse-to-fiNE alignment framework. Specifically, CONE dynamically slices the long video into candidate windows via a sliding window approach. Centering at windows, CONE (1) learns the inter-window (coarse-grained) semantic variance through contrastive learning and speeds up inference by pre-filtering the candidate windows relevant to the NL query, and (2) conducts intra-window (fine-grained) candidate moments ranking utilizing the powerful multi-modal alignment ability of the contrastive vision-text pre-trained model EgoVLP. On the blind test set, CONE achieves 15.26 and 9.24 for R1@IoU=0.3 and R1@IoU=0.5, respectively.

Fan Ma, Linchao Zhu, Yi Yang et al.

In this paper, we study an intermediate form of supervision, i.e., single-frame supervision, for temporal action localization (TAL). To obtain the single-frame supervision, the annotators are asked to identify only a single frame within the temporal window of an action. This can significantly reduce the labor cost of obtaining full supervision which requires annotating the action boundary. Compared to the weak supervision that only annotates the video-level label, the single-frame supervision introduces extra temporal action signals while maintaining low annotation overhead. To make full use of such single-frame supervision, we propose a unified system called SF-Net. First, we propose to predict an actionness score for each video frame. Along with a typical category score, the actionness score can provide comprehensive information about the occurrence of a potential action and aid the temporal boundary refinement during inference. Second, we mine pseudo action and background frames based on the single-frame annotations. We identify pseudo action frames by adaptively expanding each annotated single frame to its nearby, contextual frames and we mine pseudo background frames from all the unannotated frames across multiple videos. Together with the ground-truth labeled frames, these pseudo-labeled frames are further used for training the classifier. In extensive experiments on THUMOS14, GTEA, and BEOID, SF-Net significantly improves upon state-of-the-art weakly-supervised methods in terms of both segment localization and single-frame localization. Notably, SF-Net achieves comparable results to its fully-supervised counterpart which requires much more resource intensive annotations. The code is available at https://github.com/Flowerfan/SF-Net.

Xinyu Gong, Heng Wang, Zheng Shou et al.

Conventional video models rely on a single stream to capture the complex spatial-temporal features. Recent work on two-stream video models, such as SlowFast network and AssembleNet, prescribe separate streams to learn complementary features, and achieve stronger performance. However, manually designing both streams as well as the in-between fusion blocks is a daunting task, requiring to explore a tremendously large design space. Such manual exploration is time-consuming and often ends up with sub-optimal architectures when computational resources are limited and the exploration is insufficient. In this work, we present a pragmatic neural architecture search approach, which is able to search for two-stream video models in giant spaces efficiently. We design a multivariate search space, including 6 search variables to capture a wide variety of choices in designing two-stream models. Furthermore, we propose a progressive search procedure, by searching for the architecture of individual streams, fusion blocks, and attention blocks one after the other. We demonstrate two-stream models with significantly better performance can be automatically discovered in our design space. Our searched two-stream models, namely Auto-TSNet, consistently outperform other models on standard benchmarks. On Kinetics, compared with the SlowFast model, our Auto-TSNet-L model reduces FLOPS by nearly 11 times while achieving the same accuracy 78.9%. On Something-Something-V2, Auto-TSNet-M improves the accuracy by at least 2% over other methods which use less than 50 GFLOPS per video.

Xudong Lin, Zheng Shou, Shih-Fu Chang

Recently, Weakly-supervised Temporal Action Localization (WTAL) has been densely studied but there is still a large gap between weakly-supervised models and fully-supervised models. It is practical and intuitive to annotate temporal boundaries of a few examples and utilize them to help WTAL models better detect actions. However, the train-test discrepancy of action localization strategy prevents WTAL models from leveraging semi-supervision for further improvement. At training time, attention or multiple instance learning is used to aggregate predictions of each snippet for video-level classification; at test time, they first obtain action score sequences over time, then truncate segments of scores higher than a fixed threshold, and post-process action segments. The inconsistent strategy makes it hard to explicitly supervise the action localization model with temporal boundary annotations at training time. In this paper, we propose a Train-Test Consistent framework, TTC-Loc. In both training and testing time, our TTC-Loc localizes actions by comparing scores of action classes and predicted threshold, which enables it to be trained with semi-supervision. By fixing the train-test discrepancy, our TTC-Loc significantly outperforms the state-of-the-art performance on THUMOS'14, ActivityNet 1.2 and 1.3 when only video-level labels are provided for training. With full annotations of only one video per class and video-level labels for the other videos, our TTC-Loc further boosts the performance and achieves 33.4\% mAP (IoU threshold 0.5) on THUMOS's 14.

Jiawei Ma, Zheng Shou, Alireza Zareian et al.

Many real-world applications involve multivariate, geo-tagged time series data: at each location, multiple sensors record corresponding measurements. For example, air quality monitoring system records PM2.5, CO, etc. The resulting time-series data often has missing values due to device outages or communication errors. In order to impute the missing values, state-of-the-art methods are built on Recurrent Neural Networks (RNN), which process each time stamp sequentially, prohibiting the direct modeling of the relationship between distant time stamps. Recently, the self-attention mechanism has been proposed for sequence modeling tasks such as machine translation, significantly outperforming RNN because the relationship between each two time stamps can be modeled explicitly. In this paper, we are the first to adapt the self-attention mechanism for multivariate, geo-tagged time series data. In order to jointly capture the self-attention across multiple dimensions, including time, location and the sensor measurements, while maintain low computational complexity, we propose a novel approach called Cross-Dimensional Self-Attention (CDSA) to process each dimension sequentially, yet in an order-independent manner. Our extensive experiments on four real-world datasets, including three standard benchmarks and our newly collected NYC-traffic dataset, demonstrate that our approach outperforms the state-of-the-art imputation and forecasting methods. A detailed systematic analysis confirms the effectiveness of our design choices.

Zheng Shou, Xudong Lin, Yannis Kalantidis et al.

Motion has shown to be useful for video understanding, where motion is typically represented by optical flow. However, computing flow from video frames is very time-consuming. Recent works directly leverage the motion vectors and residuals readily available in the compressed video to represent motion at no cost. While this avoids flow computation, it also hurts accuracy since the motion vector is noisy and has substantially reduced resolution, which makes it a less discriminative motion representation. To remedy these issues, we propose a lightweight generator network, which reduces noises in motion vectors and captures fine motion details, achieving a more Discriminative Motion Cue (DMC) representation. Since optical flow is a more accurate motion representation, we train the DMC generator to approximate flow using a reconstruction loss and a generative adversarial loss, jointly with the downstream action classification task. Extensive evaluations on three action recognition benchmarks (HMDB-51, UCF-101, and a subset of Kinetics) confirm the effectiveness of our method. Our full system, consisting of the generator and the classifier, is coined as DMC-Net which obtains high accuracy close to that of using flow and runs two orders of magnitude faster than using optical flow at inference time.

Hang Gao, Zheng Shou, Alireza Zareian et al.

Deep neural networks suffer from over-fitting and catastrophic forgetting when trained with small data. One natural remedy for this problem is data augmentation, which has been recently shown to be effective. However, previous works either assume that intra-class variances can always be generalized to new classes, or employ naive generation methods to hallucinate finite examples without modeling their latent distributions. In this work, we propose Covariance-Preserving Adversarial Augmentation Networks to overcome existing limits of low-shot learning. Specifically, a novel Generative Adversarial Network is designed to model the latent distribution of each novel class given its related base counterparts. Since direct estimation of novel classes can be inductively biased, we explicitly preserve covariance information as the `variability' of base examples during the generation process. Empirical results show that our model can generate realistic yet diverse examples, leading to substantial improvements on the ImageNet benchmark over the state of the art.

Zheng Shou, Hang Gao, Lei Zhang et al.

Temporal Action Localization (TAL) in untrimmed video is important for many applications. But it is very expensive to annotate the segment-level ground truth (action class and temporal boundary). This raises the interest of addressing TAL with weak supervision, namely only video-level annotations are available during training). However, the state-of-the-art weakly-supervised TAL methods only focus on generating good Class Activation Sequence (CAS) over time but conduct simple thresholding on CAS to localize actions. In this paper, we first develop a novel weakly-supervised TAL framework called AutoLoc to directly predict the temporal boundary of each action instance. We propose a novel Outer-Inner-Contrastive (OIC) loss to automatically discover the needed segment-level supervision for training such a boundary predictor. Our method achieves dramatically improved performance: under the IoU threshold 0.5, our method improves mAP on THUMOS'14 from 13.7% to 21.2% and mAP on ActivityNet from 7.4% to 27.3%. It is also very encouraging to see that our weakly-supervised method achieves comparable results with some fully-supervised methods.

Zheng Shou, Junting Pan, Jonathan Chan et al.

We aim to tackle a novel task in action detection - Online Detection of Action Start (ODAS) in untrimmed, streaming videos. The goal of ODAS is to detect the start of an action instance, with high categorization accuracy and low detection latency. ODAS is important in many applications such as early alert generation to allow timely security or emergency response. We propose three novel methods to specifically address the challenges in training ODAS models: (1) hard negative samples generation based on Generative Adversarial Network (GAN) to distinguish ambiguous background, (2) explicitly modeling the temporal consistency between data around action start and data succeeding action start, and (3) adaptive sampling strategy to handle the scarcity of training data. We conduct extensive experiments using THUMOS'14 and ActivityNet. We show that our proposed methods lead to significant performance gains and improve the state-of-the-art methods. An ablation study confirms the effectiveness of each proposed method.

Tianwei Lin, Xu Zhao, Zheng Shou

Temporal action detection is a very important yet challenging problem, since videos in real applications are usually long, untrimmed and contain multiple action instances. This problem requires not only recognizing action categories but also detecting start time and end time of each action instance. Many state-of-the-art methods adopt the "detection by classification" framework: first do proposal, and then classify proposals. The main drawback of this framework is that the boundaries of action instance proposals have been fixed during the classification step. To address this issue, we propose a novel Single Shot Action Detector (SSAD) network based on 1D temporal convolutional layers to skip the proposal generation step via directly detecting action instances in untrimmed video. On pursuit of designing a particular SSAD network that can work effectively for temporal action detection, we empirically search for the best network architecture of SSAD due to lacking existing models that can be directly adopted. Moreover, we investigate into input feature types and fusion strategies to further improve detection accuracy. We conduct extensive experiments on two challenging datasets: THUMOS 2014 and MEXaction2. When setting Intersection-over-Union threshold to 0.5 during evaluation, SSAD significantly outperforms other state-of-the-art systems by increasing mAP from 19.0% to 24.6% on THUMOS 2014 and from 7.4% to 11.0% on MEXaction2.

Du Tran, Jamie Ray, Zheng Shou et al.

Learning image representations with ConvNets by pre-training on ImageNet has proven useful across many visual understanding tasks including object detection, semantic segmentation, and image captioning. Although any image representation can be applied to video frames, a dedicated spatiotemporal representation is still vital in order to incorporate motion patterns that cannot be captured by appearance based models alone. This paper presents an empirical ConvNet architecture search for spatiotemporal feature learning, culminating in a deep 3-dimensional (3D) Residual ConvNet. Our proposed architecture outperforms C3D by a good margin on Sports-1M, UCF101, HMDB51, THUMOS14, and ASLAN while being 2 times faster at inference time, 2 times smaller in model size, and having a more compact representation.

Tianwei Lin, Xu Zhao, Zheng Shou

In this notebook paper, we describe our approach in the submission to the temporal action proposal (task 3) and temporal action localization (task 4) of ActivityNet Challenge hosted at CVPR 2017. Since the accuracy in action classification task is already very high (nearly 90% in ActivityNet dataset), we believe that the main bottleneck for temporal action localization is the quality of action proposals. Therefore, we mainly focus on the temporal action proposal task and propose a new proposal model based on temporal convolutional network. Our approach achieves the state-of-the-art performances on both temporal action proposal task and temporal action localization task.

Zheng Shou, Jonathan Chan, Alireza Zareian et al.

Temporal action localization is an important yet challenging problem. Given a long, untrimmed video consisting of multiple action instances and complex background contents, we need not only to recognize their action categories, but also to localize the start time and end time of each instance. Many state-of-the-art systems use segment-level classifiers to select and rank proposal segments of pre-determined boundaries. However, a desirable model should move beyond segment-level and make dense predictions at a fine granularity in time to determine precise temporal boundaries. To this end, we design a novel Convolutional-De-Convolutional (CDC) network that places CDC filters on top of 3D ConvNets, which have been shown to be effective for abstracting action semantics but reduce the temporal length of the input data. The proposed CDC filter performs the required temporal upsampling and spatial downsampling operations simultaneously to predict actions at the frame-level granularity. It is unique in jointly modeling action semantics in space-time and fine-grained temporal dynamics. We train the CDC network in an end-to-end manner efficiently. Our model not only achieves superior performance in detecting actions in every frame, but also significantly boosts the precision of localizing temporal boundaries. Finally, the CDC network demonstrates a very high efficiency with the ability to process 500 frames per second on a single GPU server. We will update the camera-ready version and publish the source codes online soon.

Dongang Wang, Zheng Shou, Hongyi Liu et al.

EventNet is a large-scale video corpus and event ontology consisting of 500 events associated with event-specific concepts. In order to improve the quality of the current EventNet, we conduct the following steps and introduce EventNet version 1.1: (1) manually verify the correctness of event labels for all videos; (2) remove the YouTube user bias by limiting the maximum number of videos in each event from the same YouTube user as 3; (3) remove the videos which are currently not accessible online; (4) remove the video belonging to multiple event categories. After the above procedure, some events may contain only a small number of videos, and therefore we crawl more videos for those events to ensure every event will contain more than 50 videos. Finally, EventNet version 1.1 contains 67,641 videos, 500 events, and 5,028 event-specific concepts. In addition, we train a Convolutional Neural Network (CNN) model for event classification via fine-tuning AlexNet using EventNet version 1.1. Then we use the trained CNN model to extract FC7 layer feature and train binary classifiers using linear SVM for each event-specific concept. We believe this new version of EventNet will significantly facilitate research in computer vision and multimedia, and will put it online for public downloading in the future.

Zheng Shou, Dongang Wang, Shih-Fu Chang

We address temporal action localization in untrimmed long videos. This is important because videos in real applications are usually unconstrained and contain multiple action instances plus video content of background scenes or other activities. To address this challenging issue, we exploit the effectiveness of deep networks in temporal action localization via three segment-based 3D ConvNets: (1) a proposal network identifies candidate segments in a long video that may contain actions; (2) a classification network learns one-vs-all action classification model to serve as initialization for the localization network; and (3) a localization network fine-tunes on the learned classification network to localize each action instance. We propose a novel loss function for the localization network to explicitly consider temporal overlap and therefore achieve high temporal localization accuracy. Only the proposal network and the localization network are used during prediction. On two large-scale benchmarks, our approach achieves significantly superior performances compared with other state-of-the-art systems: mAP increases from 1.7% to 7.4% on MEXaction2 and increases from 15.0% to 19.0% on THUMOS 2014, when the overlap threshold for evaluation is set to 0.5.