Prior work has studied different visual modalities in isolation and developed separate architectures for recognition of images, videos, and 3D data. Instead, in this paper, we propose a single model which excels at classifying images, videos, and single-view 3D data using exactly the same model parameters. Our 'Omnivore' model leverages the flexibility of transformer-based architectures and is trained jointly on classification tasks from different modalities. Omnivore is simple to train, uses off-the-shelf standard datasets, and performs at-par or better than modality-specific models of the same size. A single Omnivore model obtains 86.0% on ImageNet, 84.1% on Kinetics, and 67.1% on SUN RGB-D. After finetuning, our models outperform prior work on a variety of vision tasks and generalize across modalities. Omnivore's shared visual representation naturally enables cross-modal recognition without access to correspondences between modalities. We hope our results motivate researchers to model visual modalities together.
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.
Transformer models have shown great success handling long-range interactions, making them a promising tool for modeling video. However, they lack inductive biases and scale quadratically with input length. These limitations are further exacerbated when dealing with the high dimensionality introduced by the temporal dimension. While there are surveys analyzing the advances of Transformers for vision, none focus on an in-depth analysis of video-specific designs. In this survey, we analyze the main contributions and trends of works leveraging Transformers to model video. Specifically, we delve into how videos are handled at the input level first. Then, we study the architectural changes made to deal with video more efficiently, reduce redundancy, re-introduce useful inductive biases, and capture long-term temporal dynamics. In addition, we provide an overview of different training regimes and explore effective self-supervised learning strategies for video. Finally, we conduct a performance comparison on the most common benchmark for Video Transformers (i.e., action classification), finding them to outperform 3D ConvNets even with less computational complexity.
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.
This paper presents the baseline method proposed for the Sports Video task part of the MediaEval 2021 benchmark. This task proposes a stroke detection and a stroke classification subtasks. This baseline addresses both subtasks. The spatio-temporal CNN architecture and the training process of the model are tailored according to the addressed subtask. The method has the purpose of helping the participants to solve the task and is not meant to reach stateof-the-art performance. Still, for the detection task, the baseline is performing better than the other participants, which stresses the difficulty of such a task.
We present Masked Feature Prediction (MaskFeat) for self-supervised pre-training of video models. Our approach first randomly masks out a portion of the input sequence and then predicts the feature of the masked regions. We study five different types of features and find Histograms of Oriented Gradients (HOG), a hand-crafted feature descriptor, works particularly well in terms of both performance and efficiency. We observe that the local contrast normalization in HOG is essential for good results, which is in line with earlier work using HOG for visual recognition. Our approach can learn abundant visual knowledge and drive large-scale Transformer-based models. Without using extra model weights or supervision, MaskFeat pre-trained on unlabeled videos achieves unprecedented results of 86.7% with MViT-L on Kinetics-400, 88.3% on Kinetics-600, 80.4% on Kinetics-700, 39.8 mAP on AVA, and 75.0% on SSv2. MaskFeat further generalizes to image input, which can be interpreted as a video with a single frame and obtains competitive results on ImageNet.
In learning action recognition, models are typically pre-trained on object recognition with images, such as ImageNet, and later fine-tuned on target action recognition with videos. This approach has achieved good empirical performance especially with recent transformer-based video architectures. While recently many works aim to design more advanced transformer architectures for action recognition, less effort has been made on how to train video transformers. In this work, we explore several training paradigms and present two findings. First, video transformers benefit from joint training on diverse video datasets and label spaces (e.g., Kinetics is appearance-focused while SomethingSomething is motion-focused). Second, by further co-training with images (as single-frame videos), the video transformers learn even better video representations. We term this approach as Co-training Videos and Images for Action Recognition (CoVeR). In particular, when pretrained on ImageNet-21K based on the TimeSFormer architecture, CoVeR improves Kinetics-400 Top-1 Accuracy by 2.4%, Kinetics-600 by 2.3%, and SomethingSomething-v2 by 2.3%. When pretrained on larger-scale image datasets following previous state-of-the-art, CoVeR achieves best results on Kinetics-400 (87.2%), Kinetics-600 (87.9%), Kinetics-700 (79.8%), SomethingSomething-v2 (70.9%), and Moments-in-Time (46.1%), with a simple spatio-temporal video transformer.
Videos are a rich source for self-supervised learning (SSL) of visual representations due to the presence of natural temporal transformations of objects. However, current methods typically randomly sample video clips for learning, which results in an imperfect supervisory signal. In this work, we propose PreViTS, an SSL framework that utilizes an unsupervised tracking signal for selecting clips containing the same object, which helps better utilize temporal transformations of objects. PreViTS further uses the tracking signal to spatially constrain the frame regions to learn from and trains the model to locate meaningful objects by providing supervision on Grad-CAM attention maps. To evaluate our approach, we train a momentum contrastive (MoCo) encoder on VGG-Sound and Kinetics-400 datasets with PreViTS. Training with PreViTS outperforms representations learnt by contrastive strategy alone on video downstream tasks, obtaining state-of-the-art performance on action classification. PreViTS helps learn feature representations that are more robust to changes in background and context, as seen by experiments on datasets with background changes. Learning from large-scale videos with PreViTS could lead to more accurate and robust visual feature representations.
In this paper, we study Multiscale Vision Transformers (MViTv2) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTv2s' pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViTv2 has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 58.7 boxAP on COCO object detection as well as 86.1% on Kinetics-400 video classification. Code and models are available at https://github.com/facebookresearch/mvit.
In this paper, we propose self-supervised training for video transformers using unlabeled video data. From a given video, we create local and global spatiotemporal views with varying spatial sizes and frame rates. Our self-supervised objective seeks to match the features of these different views representing the same video, to be invariant to spatiotemporal variations in actions. To the best of our knowledge, the proposed approach is the first to alleviate the dependency on negative samples or dedicated memory banks in Self-supervised Video Transformer (SVT). Further, owing to the flexibility of Transformer models, SVT supports slow-fast video processing within a single architecture using dynamically adjusted positional encoding and supports long-term relationship modeling along spatiotemporal dimensions. Our approach performs well on four action recognition benchmarks (Kinetics-400, UCF-101, HMDB-51, and SSv2) and converges faster with small batch sizes. Code: https://git.io/J1juJ
Models of human motion commonly focus either on trajectory prediction or action classification but rarely both. The marked heterogeneity and intricate compositionality of human motion render each task vulnerable to the data degradation and distributional shift common to real-world scenarios. A sufficiently expressive generative model of action could in theory enable data conditioning and distributional resilience within a unified framework applicable to both tasks. Here we propose a novel architecture based on hierarchical variational autoencoders and deep graph convolutional neural networks for generating a holistic model of action over multiple time-scales. We show this Hierarchical Graph-convolutional Variational Autoencoder (HG-VAE) to be capable of generating coherent actions, detecting out-of-distribution data, and imputing missing data by gradient ascent on the model's posterior. Trained and evaluated on H3.6M and the largest collection of open source human motion data, AMASS, we show HG-VAE can facilitate downstream discriminative learning better than baseline models.
Automated visual understanding of our diverse and open world demands computer vision models to generalize well with minimal customization for specific tasks, similar to human vision. Computer vision foundation models, which are trained on diverse, large-scale dataset and can be adapted to a wide range of downstream tasks, are critical for this mission to solve real-world computer vision applications. While existing vision foundation models such as CLIP, ALIGN, and Wu Dao 2.0 focus mainly on mapping images and textual representations to a cross-modal shared representation, we introduce a new computer vision foundation model, Florence, to expand the representations from coarse (scene) to fine (object), from static (images) to dynamic (videos), and from RGB to multiple modalities (caption, depth). By incorporating universal visual-language representations from Web-scale image-text data, our Florence model can be easily adapted for various computer vision tasks, such as classification, retrieval, object detection, VQA, image caption, video retrieval and action recognition. Moreover, Florence demonstrates outstanding performance in many types of transfer learning: fully sampled fine-tuning, linear probing, few-shot transfer and zero-shot transfer for novel images and objects. All of these properties are critical for our vision foundation model to serve general purpose vision tasks. Florence achieves new state-of-the-art results in majority of 44 representative benchmarks, e.g., ImageNet-1K zero-shot classification with top-1 accuracy of 83.74 and the top-5 accuracy of 97.18, 62.4 mAP on COCO fine tuning, 80.36 on VQA, and 87.8 on Kinetics-600.
Large-scale NLP models have been shown to significantly improve the performance on language tasks with no signs of saturation. They also demonstrate amazing few-shot capabilities like that of human beings. This paper aims to explore large-scale models in computer vision. We tackle three major issues in training and application of large vision models, including training instability, resolution gaps between pre-training and fine-tuning, and hunger on labelled data. Three main techniques are proposed: 1) a residual-post-norm method combined with cosine attention to improve training stability; 2) A log-spaced continuous position bias method to effectively transfer models pre-trained using low-resolution images to downstream tasks with high-resolution inputs; 3) A self-supervised pre-training method, SimMIM, to reduce the needs of vast labeled images. Through these techniques, this paper successfully trained a 3 billion-parameter Swin Transformer V2 model, which is the largest dense vision model to date, and makes it capable of training with images of up to 1,536$\times$1,536 resolution. It set new performance records on 4 representative vision tasks, including ImageNet-V2 image classification, COCO object detection, ADE20K semantic segmentation, and Kinetics-400 video action classification. Also note our training is much more efficient than that in Google's billion-level visual models, which consumes 40 times less labelled data and 40 times less training time. Code is available at \url{https://github.com/microsoft/Swin-Transformer}.
Video-based computer vision tasks can benefit from estimation of the salient regions and interactions between those regions. Traditionally, this has been done by identifying the object regions in the images by utilizing pre-trained models to perform object detection, object segmentation and/or object pose estimation. Although using pre-trained models is a viable approach, it has several limitations in the need for an exhaustive annotation of object categories, a possible domain gap between datasets, and a bias that is typically present in pre-trained models. In this work, we propose to utilize the common rationale that a sequence of video frames capture a set of common objects and interactions between them, thus a notion of co-segmentation between the video frame features may equip the model with the ability to automatically focus on task-specific salient regions and improve the underlying task's performance in an end-to-end manner. In this regard, we propose a generic module called ``Co-Segmentation inspired Attention Module'' (COSAM) that can be plugged in to any CNN model to promote the notion of co-segmentation based attention among a sequence of video frame features. We show the application of COSAM in three video-based tasks namely: 1) Video-based person re-ID, 2) Video captioning, & 3) Video action classification and demonstrate that COSAM is able to capture the task-specific salient regions in video frames, thus leading to notable performance improvements along with interpretable attention maps for a variety of video-based vision tasks, with possible application to other video-based vision tasks as well.
Recognizing human actions is fundamentally a spatio-temporal reasoning problem, and should be, at least to some extent, invariant to the appearance of the human and the objects involved. Motivated by this hypothesis, in this work, we take an object-centric approach to action recognition. Multiple works have studied this setting before, yet it remains unclear (i) how well a carefully crafted, spatio-temporal layout-based method can recognize human actions, and (ii) how, and when, to fuse the information from layout and appearance-based models. The main focus of this paper is compositional/few-shot action recognition, where we advocate the usage of multi-head attention (proven to be effective for spatial reasoning) over spatio-temporal layouts, i.e., configurations of object bounding boxes. We evaluate different schemes to inject video appearance information to the system, and benchmark our approach on background cluttered action recognition. On the Something-Else and Action Genome datasets, we demonstrate (i) how to extend multi-head attention for spatio-temporal layout-based action recognition, (ii) how to improve the performance of appearance-based models by fusion with layout-based models, (iii) that even on non-compositional background-cluttered video datasets, a fusion between layout- and appearance-based models improves the performance.
Deep learning has shown remarkable progress in a wide range of problems. However, efficient training of such models requires large-scale datasets, and getting annotations for such datasets can be challenging and costly. In this work, we explore the use of user-generated freely available labels from web videos for video understanding. We create a benchmark dataset consisting of around 2 million videos with associated user-generated annotations and other meta information. We utilize the collected dataset for action classification and demonstrate its usefulness with existing small-scale annotated datasets, UCF101 and HMDB51. We study different loss functions and two pretraining strategies, simple and self-supervised learning. We also show how a network pretrained on the proposed dataset can help against video corruption and label noise in downstream datasets. We present this as a benchmark dataset in noisy learning for video understanding. The dataset, code, and trained models will be publicly available for future research.
Spatial convolutions are widely used in numerous deep video models. It fundamentally assumes spatio-temporal invariance, i.e., using shared weights for every location in different frames. This work presents Temporally-Adaptive Convolutions (TAdaConv) for video understanding, which shows that adaptive weight calibration along the temporal dimension is an efficient way to facilitate modelling complex temporal dynamics in videos. Specifically, TAdaConv empowers the spatial convolutions with temporal modelling abilities by calibrating the convolution weights for each frame according to its local and global temporal context. Compared to previous temporal modelling operations, TAdaConv is more efficient as it operates over the convolution kernels instead of the features, whose dimension is an order of magnitude smaller than the spatial resolutions. Further, the kernel calibration brings an increased model capacity. We construct TAda2D and TAdaConvNeXt networks by replacing the 2D convolutions in ResNet and ConvNeXt with TAdaConv, which leads to at least on par or better performance compared to state-of-the-art approaches on multiple video action recognition and localization benchmarks. We also demonstrate that as a readily plug-in operation with negligible computation overhead, TAdaConv can effectively improve many existing video models with a convincing margin.
This paper proposes a fusion method of modalities extracted from video through a three-stream network with spatio-temporal and temporal convolutions for fine-grained action classification in sport. It is applied to TTStroke-21 dataset which consists of untrimmed videos of table tennis games. The goal is to detect and classify table tennis strokes in the videos, the first step of a bigger scheme aiming at giving feedback to the players for improving their performance. The three modalities are raw RGB data, the computed optical flow and the estimated pose of the player. The network consists of three branches with attention blocks. Features are fused at the latest stage of the network using bilinear layers. Compared to previous approaches, the use of three modalities allows faster convergence and better performances on both tasks: classification of strokes with known temporal boundaries and joint segmentation and classification. The pose is also further investigated in order to offer richer feedback to the athletes.
Most recent view-invariant action recognition and performance assessment approaches rely on a large amount of annotated 3D skeleton data to extract view-invariant features. However, acquiring 3D skeleton data can be cumbersome, if not impractical, in in-the-wild scenarios. To overcome this problem, we present a novel unsupervised approach that learns to extract view-invariant 3D human pose representation from a 2D image without using 3D joint data. Our model is trained by exploiting the intrinsic view-invariant properties of human pose between simultaneous frames from different viewpoints and their equivariant properties between augmented frames from the same viewpoint. We evaluate the learned view-invariant pose representations for two downstream tasks. We perform comparative experiments that show improvements on the state-of-the-art unsupervised cross-view action classification accuracy on NTU RGB+D by a significant margin, on both RGB and depth images. We also show the efficiency of transferring the learned representations from NTU RGB+D to obtain the first ever unsupervised cross-view and cross-subject rank correlation results on the multi-view human movement quality dataset, QMAR, and marginally improve on the-state-of-the-art supervised results for this dataset. We also carry out ablation studies to examine the contributions of the different components of our proposed network.