Despite the recent success of video self-supervised learning models, there is much still to be understood about their generalization capability. In this paper, we investigate how sensitive video self-supervised learning is to the current conventional benchmark and whether methods generalize beyond the canonical evaluation setting. We do this across four different factors of sensitivity: domain, samples, actions and task. Our study which encompasses over 500 experiments on 7 video datasets, 9 self-supervised methods and 6 video understanding tasks, reveals that current benchmarks in video self-supervised learning are not good indicators of generalization along these sensitivity factors. Further, we find that self-supervised methods considerably lag behind vanilla supervised pre-training, especially when domain shift is large and the amount of available downstream samples are low. From our analysis, we distill the SEVERE-benchmark, a subset of our experiments, and discuss its implication for evaluating the generalizability of representations obtained by existing and future self-supervised video learning methods.
Large-scale pretrained image-text models have shown incredible zero-shot performance in a handful of tasks, including video ones such as action recognition and text-to-video retrieval. However, these models have not been adapted to video, mainly because they do not account for the time dimension but also because video frames are different from the typical images (e.g., containing motion blur, and less sharpness). In this paper, we present a fine-tuning strategy to refine these large-scale pretrained image-text models for zero-shot video understanding tasks. We show that by carefully adapting these models we obtain considerable improvements on two zero-shot Action Recognition tasks and three zero-shot Text-to-video Retrieval tasks. The code is available at https://github.com/bryant1410/fitclip
Pre-training video transformers on extra large-scale datasets is generally required to achieve premier performance on relatively small datasets. In this paper, we show that video masked autoencoders (VideoMAE) are data-efficient learners for self-supervised video pre-training (SSVP). We are inspired by the recent ImageMAE and propose customized video tube masking with an extremely high ratio. This simple design makes video reconstruction a more challenging self-supervision task, thus encouraging extracting more effective video representations during this pre-training process. We obtain three important findings on SSVP: (1) An extremely high proportion of masking ratio (i.e., 90% to 95%) still yields favorable performance of VideoMAE. The temporally redundant video content enables a higher masking ratio than that of images. (2) VideoMAE achieves impressive results on very small datasets (i.e., around 3k-4k videos) without using any extra data. (3) VideoMAE shows that data quality is more important than data quantity for SSVP. Domain shift between pre-training and target datasets is an important issue. Notably, our VideoMAE with the vanilla ViT can achieve 87.4% on Kinetics-400, 75.4% on Something-Something V2, 91.3% on UCF101, and 62.6% on HMDB51, without using any extra data. Code is available at https://github.com/MCG-NJU/VideoMAE.
Batch Normalization's (BN) unique property of depending on other samples in a batch is known to cause problems in several tasks, including sequence modeling. Yet, BN-related issues are hardly studied for long video understanding, despite the ubiquitous use of BN in CNNs (Convolutional Neural Networks) for feature extraction. Especially in surgical workflow analysis, where the lack of pretrained feature extractors has led to complex, multi-stage training pipelines, limited awareness of BN issues may have hidden the benefits of training CNNs and temporal models end to end. In this paper, we analyze pitfalls of BN in video learning, including issues specific to online tasks such as a 'cheating' effect in anticipation. We observe that BN's properties create major obstacles for end-to-end learning. However, using BN-free backbones, even simple CNN-LSTMs beat the state of the art {\color{\colorrevtwo}on three surgical workflow benchmarks} by utilizing adequate end-to-end training strategies which maximize temporal context. We conclude that awareness of BN's pitfalls is crucial for effective end-to-end learning in surgical tasks. By reproducing results on natural-video datasets, we hope our insights will benefit other areas of video learning as well. Code is available at: \url{https://gitlab.com/nct_tso_public/pitfalls_bn}
Modeling and automatically recognizing surgical activities are fundamental steps toward automation in surgery and play important roles in providing timely feedback to surgeons. Accurately recognizing surgical activities in video poses a challenging problem that requires an effective means of learning both spatial and temporal dynamics. Human gaze and visual saliency carry important information about visual attention and can be used to extract more relevant features that better reflect these spatial and temporal dynamics. In this study, we propose to use human gaze with a spatio-temporal attention mechanism for activity recognition in surgical videos. Our model consists of an I3D-based architecture, learns spatio-temporal features using 3D convolutions, as well as learns an attention map using human gaze as supervision. We evaluate our model on the Suturing task of JIGSAWS which is a publicly available surgical video understanding dataset. To our knowledge, we are the first to use human gaze for surgical activity recognition. Our results and ablation studies support the contribution of using human gaze to guide attention by outperforming state-of-the art models with an accuracy of 85.4%.
Query-based video grounding is an important yet challenging task in video understanding, which aims to localize the target segment in an untrimmed video according to a sentence query. Most previous works achieve significant progress by addressing this task in a fully-supervised manner with segment-level labels, which require high labeling cost. Although some recent efforts develop weakly-supervised methods that only need the video-level knowledge, they generally match multiple pre-defined segment proposals with query and select the best one, which lacks fine-grained frame-level details for distinguishing frames with high repeatability and similarity within the entire video. To alleviate the above limitations, we propose a self-contrastive learning framework to address the query-based video grounding task under a weakly-supervised setting. Firstly, instead of utilizing redundant segment proposals, we propose a new grounding scheme that learns frame-wise matching scores referring to the query semantic to predict the possible foreground frames by only using the video-level annotations. Secondly, since some predicted frames (i.e., boundary frames) are relatively coarse and exhibit similar appearance to their adjacent frames, we propose a coarse-to-fine contrastive learning paradigm to learn more discriminative frame-wise representations for distinguishing the false positive frames. In particular, we iteratively explore multi-scale hard negative samples that are close to positive samples in the representation space for distinguishing fine-grained frame-wise details, thus enforcing more accurate segment grounding. Extensive experiments on two challenging benchmarks demonstrate the superiority of our proposed method compared with the state-of-the-art methods.
Dynamic scene graph generation from a video is challenging due to the temporal dynamics of the scene and the inherent temporal fluctuations of predictions. We hypothesize that capturing long-term temporal dependencies is the key to effective generation of dynamic scene graphs. We propose to learn the long-term dependencies in a video by capturing the object-level consistency and inter-object relationship dynamics over object-level long-term tracklets using transformers. Experimental results demonstrate that our Dynamic Scene Graph Detection Transformer (DSG-DETR) outperforms state-of-the-art methods by a significant margin on the benchmark dataset Action Genome. Our ablation studies validate the effectiveness of each component of the proposed approach. The source code is available at https://github.com/Shengyu-Feng/DSG-DETR.
Generic Boundary Detection (GBD) aims at locating the general boundaries that divide videos into semantically coherent and taxonomy-free units, and could serve as an important pre-processing step for long-form video understanding. Previous works often separately handle these different types of generic boundaries with specific designs of deep networks from simple CNN to LSTM. Instead, in this paper, we present Temporal Perceiver, a general architecture with Transformer, offering a unified solution to the detection of arbitrary generic boundaries, ranging from shot-level, event-level, to scene-level GBDs. The core design is to introduce a small set of latent feature queries as anchors to compress the redundant video input into a fixed dimension via cross-attention blocks. Thanks to this fixed number of latent units, it greatly reduces the quadratic complexity of attention operation to a linear form of input frames. Specifically, to explicitly leverage the temporal structure of videos, we construct two types of latent feature queries: boundary queries and context queries, which handle the semantic incoherence and coherence accordingly. Moreover, to guide the learning of latent feature queries, we propose an alignment loss on the cross-attention maps to explicitly encourage the boundary queries to attend on the top boundary candidates. Finally, we present a sparse detection head on the compressed representation, and directly output the final boundary detection results without any post-processing module. We test our Temporal Perceiver on a variety of GBD benchmarks. Our method obtains the state-of-the-art results on all benchmarks with RGB single-stream features: SoccerNet-v2 (81.9% avg-mAP), Kinetics-GEBD (86.0% avg-f1), TAPOS (73.2% avg-f1), MovieScenes (51.9% AP and 53.1% Miou) and MovieNet (53.3% AP and 53.2% Miou), demonstrating the generalization ability of our Temporal Perceiver.
We constantly integrate our knowledge and understanding of the world to enhance our interpretation of what we see. This ability is crucial in application domains which entail reasoning about multiple entities and concepts, such as AI-augmented surgery. In this paper, we propose a novel way of integrating conceptual knowledge into temporal analysis tasks via temporal concept graph networks. In the proposed networks, a global knowledge graph is incorporated into the temporal analysis of surgical instances, learning the meaning of concepts and relations as they apply to the data. We demonstrate our results in surgical video data for tasks such as verification of critical view of safety, as well as estimation of Parkland grading scale. The results show that our method improves the recognition and detection of complex benchmarks as well as enables other analytic applications of interest.
Maintaining proper form while exercising is important for preventing injuries and maximizing muscle mass gains. Detecting errors in workout form naturally requires estimating human's body pose. However, off-the-shelf pose estimators struggle to perform well on the videos recorded in gym scenarios due to factors such as camera angles, occlusion from gym equipment, illumination, and clothing. To aggravate the problem, the errors to be detected in the workouts are very subtle. To that end, we propose to learn exercise-oriented image and video representations from unlabeled samples such that a small dataset annotated by experts suffices for supervised error detection. In particular, our domain knowledge-informed self-supervised approaches (pose contrastive learning and motion disentangling) exploit the harmonic motion of the exercise actions, and capitalize on the large variances in camera angles, clothes, and illumination to learn powerful representations. To facilitate our self-supervised pretraining, and supervised finetuning, we curated a new exercise dataset, \emph{Fitness-AQA} (\url{https://github.com/ParitoshParmar/Fitness-AQA}), comprising of three exercises: BackSquat, BarbellRow, and OverheadPress. It has been annotated by expert trainers for multiple crucial and typically occurring exercise errors. Experimental results show that our self-supervised representations outperform off-the-shelf 2D- and 3D-pose estimators and several other baselines. We also show that our approaches can be applied to other domains/tasks such as pose estimation and dive quality assessment.
There have been many attempts to build multimodal dialog systems that can respond to a question about given audio-visual information, and the representative task for such systems is the Audio Visual Scene-Aware Dialog (AVSD). Most conventional AVSD models adopt the Convolutional Neural Network (CNN)-based video feature extractor to understand visual information. While a CNN tends to obtain both temporally and spatially local information, global information is also crucial for boosting video understanding because AVSD requires long-term temporal visual dependency and whole visual information. In this study, we apply the Transformer-based video feature that can capture both temporally and spatially global representations more efficiently than the CNN-based feature. Our AVSD model with its Transformer-based feature attains higher objective performance scores for answer generation. In addition, our model achieves a subjective score close to that of human answers in DSTC10. We observed that the Transformer-based visual feature is beneficial for the AVSD task because our model tends to correctly answer the questions that need a temporally and spatially broad range of visual information.
Self-attention based Transformer models have demonstrated impressive results for image classification and object detection, and more recently for video understanding. Inspired by this success, we investigate the application of Transformer networks for temporal action localization in videos. To this end, we present ActionFormer -- a simple yet powerful model to identify actions in time and recognize their categories in a single shot, without using action proposals or relying on pre-defined anchor windows. ActionFormer combines a multiscale feature representation with local self-attention, and uses a light-weighted decoder to classify every moment in time and estimate the corresponding action boundaries. We show that this orchestrated design results in major improvements upon prior works. Without bells and whistles, ActionFormer achieves 71.0% mAP at tIoU=0.5 on THUMOS14, outperforming the best prior model by 14.1 absolute percentage points. Further, ActionFormer demonstrates strong results on ActivityNet 1.3 (36.6% average mAP) and EPIC-Kitchens 100 (+13.5% average mAP over prior works). Our code is available at http://github.com/happyharrycn/actionformer_release.
Estimating per-pixel motion between video frames, known as optical flow, is a long-standing problem in video understanding and analysis. Most contemporary optical flow techniques largely focus on addressing the cross-image matching with feature similarity, with few methods considering how to explicitly reason over the given scene for achieving a holistic motion understanding. In this work, taking a fresh perspective, we introduce a novel graph-based approach, called adaptive graph reasoning for optical flow (AGFlow), to emphasize the value of scene/context information in optical flow. Our key idea is to decouple the context reasoning from the matching procedure, and exploit scene information to effectively assist motion estimation by learning to reason over the adaptive graph. The proposed AGFlow can effectively exploit the context information and incorporate it within the matching procedure, producing more robust and accurate results. On both Sintel clean and final passes, our AGFlow achieves the best accuracy with EPE of 1.43 and 2.47 pixels, outperforming state-of-the-art approaches by 11.2% and 13.6%, respectively.
Video compression is indispensable to most video analysis systems. Despite saving transportation bandwidth, it also deteriorates downstream video understanding tasks, especially at low-bitrate settings. To systematically investigate this problem, we first thoroughly review the previous methods, revealing that three principles, i.e., task-decoupled, label-free, and data-emerged semantic prior, are critical to a machine-friendly coding framework but are not fully satisfied so far. In this paper, we propose a traditional-neural mixed coding framework that simultaneously fulfills all these principles, by taking advantage of both traditional codecs and neural networks (NNs). On one hand, the traditional codecs can efficiently encode the pixel signal of videos but may distort the semantic information. On the other hand, highly non-linear NNs are proficient in condensing video semantics into a compact representation. The framework is optimized by ensuring that a transportation-efficient semantic representation of the video is preserved w.r.t. the coding procedure, which is spontaneously learned from unlabeled data in a self-supervised manner. The videos collaboratively decoded from two streams (codec and NN) are of rich semantics, as well as visually photo-realistic, empirically boosting several mainstream downstream video analysis task performances without any post-adaptation procedure. Furthermore, by introducing the attention mechanism and adaptive modeling scheme, the video semantic modeling ability of our approach is further enhanced. Finally, we build a low-bitrate video understanding benchmark with three downstream tasks on eight datasets, demonstrating the notable superiority of our approach. All codes, data, and models will be available at \url{https://github.com/tianyuan168326/VCS-Pytorch}.
This paper introduces a new challenge and datasets to foster research toward designing systems that can understand medical videos and provide visual answers to natural language questions. We believe medical videos may provide the best possible answers to many first aids, medical emergency, and medical education questions. Toward this, we created the MedVidCL and MedVidQA datasets and introduce the tasks of Medical Video Classification (MVC) and Medical Visual Answer Localization (MVAL), two tasks that focus on cross-modal (medical language and medical video) understanding. The proposed tasks and datasets have the potential to support the development of sophisticated downstream applications that can benefit the public and medical practitioners. Our datasets consist of 6,117 annotated videos for the MVC task and 3,010 annotated questions and answers timestamps from 899 videos for the MVAL task. These datasets have been verified and corrected by medical informatics experts. We have also benchmarked each task with the created MedVidCL and MedVidQA datasets and proposed the multimodal learning methods that set competitive baselines for future research.
We address the problem of capturing temporal information for video classification in 2D networks, without increasing their computational cost. Existing approaches focus on modifying the architecture of 2D networks (e.g. by including filters in the temporal dimension to turn them into 3D networks, or using optical flow, etc.), which increases computation cost. Instead, we propose a novel sampling strategy, where we re-order the channels of the input video, to capture short-term frame-to-frame changes. We observe that without bells and whistles, the proposed sampling strategy improves performance on multiple architectures (e.g. TSN, TRN, TSM, and MVFNet) and datasets (CATER, Something-Something-V1 and V2), up to 24% over the baseline of using the standard video input. In addition, our sampling strategies do not require training from scratch and do not increase the computational cost of training and testing. Given the generality of the results and the flexibility of the approach, we hope this can be widely useful to the video understanding community. Code is available on our website: https://github.com/kiyoon/channel_sampling.
Recent video and language pretraining frameworks lack the ability to generate sentences. We present Multimodal Video Generative Pretraining (MV-GPT), a new pretraining framework for learning from unlabelled videos which can be effectively used for generative tasks such as multimodal video captioning. Unlike recent video-language pretraining frameworks, our framework trains both a multimodal video encoder and a sentence decoder jointly. To overcome the lack of captions in unlabelled videos, we leverage the future utterance as an additional text source and propose a bidirectional generation objective -- we generate future utterances given the present mulitmodal context, and also the present utterance given future observations. With this objective, we train an encoder-decoder model end-to-end to generate a caption from raw pixels and transcribed speech directly. Our model achieves state-of-the-art performance for multimodal video captioning on four standard benchmarks, as well as for other video understanding tasks such as VideoQA, video retrieval and action classification.
Video understanding requires reasoning at multiple spatiotemporal resolutions -- from short fine-grained motions to events taking place over longer durations. Although transformer architectures have recently advanced the state-of-the-art, they have not explicitly modelled different spatiotemporal resolutions. To this end, we present Multiview Transformers for Video Recognition (MTV). Our model consists of separate encoders to represent different views of the input video with lateral connections to fuse information across views. We present thorough ablation studies of our model and show that MTV consistently performs better than single-view counterparts in terms of accuracy and computational cost across a range of model sizes. Furthermore, we achieve state-of-the-art results on six standard datasets, and improve even further with large-scale pretraining. Code and checkpoints are available at: https://github.com/google-research/scenic/tree/main/scenic/projects/mtv.
As humans, we navigate a multimodal world, building a holistic understanding from all our senses. We introduce MERLOT Reserve, a model that represents videos jointly over time -- through a new training objective that learns from audio, subtitles, and video frames. Given a video, we replace snippets of text and audio with a MASK token; the model learns by choosing the correct masked-out snippet. Our objective learns faster than alternatives, and performs well at scale: we pretrain on 20 million YouTube videos. Empirical results show that MERLOT Reserve learns strong multimodal representations. When finetuned, it sets state-of-the-art on Visual Commonsense Reasoning (VCR), TVQA, and Kinetics-600; outperforming prior work by 5%, 7%, and 1.5% respectively. Ablations show that these tasks benefit from audio pretraining -- even VCR, a QA task centered around images (without sound). Moreover, our objective enables out-of-the-box prediction, revealing strong multimodal commonsense understanding. In a fully zero-shot setting, our model obtains competitive results on four video tasks, even outperforming supervised approaches on the recently proposed Situated Reasoning (STAR) benchmark. We analyze why audio enables better vision-language representations, suggesting significant opportunities for future research. We conclude by discussing ethical and societal implications of multimodal pretraining.
Temporal sentence grounding (TSG) is crucial and fundamental for video understanding. Although the existing methods train well-designed deep networks with a large amount of data, we find that they can easily forget the rarely appeared cases in the training stage due to the off-balance data distribution, which influences the model generalization and leads to undesirable performance. To tackle this issue, we propose a memory-augmented network, called Memory-Guided Semantic Learning Network (MGSL-Net), that learns and memorizes the rarely appeared content in TSG tasks. Specifically, MGSL-Net consists of three main parts: a cross-modal inter-action module, a memory augmentation module, and a heterogeneous attention module. We first align the given video-query pair by a cross-modal graph convolutional network, and then utilize a memory module to record the cross-modal shared semantic features in the domain-specific persistent memory. During training, the memory slots are dynamically associated with both common and rare cases, alleviating the forgetting issue. In testing, the rare cases can thus be enhanced by retrieving the stored memories, resulting in better generalization. At last, the heterogeneous attention module is utilized to integrate the enhanced multi-modal features in both video and query domains. Experimental results on three benchmarks show the superiority of our method on both effectiveness and efficiency, which substantially improves the accuracy not only on the entire dataset but also on rare cases.