Transformer [70]has brought significant progress in natural language processing [17,7,54]. The vision transformer [20] also improves a series of computer vision tasks including image classification [66,88], object detection [8,37], semantic segmentation [80], object tracking [13,16], and video recognition [6,3].

In video transformers, the time dimension is often treated in the same way as the two spatial dimensions. However, in a scene where objects or the camera may move, a physical point imaged at one location in frame $t$ may be entirely unrelated to what is found at that location in frame $t+k$. These temporal correspondences should be modeled to facilitate learning about dynamic scenes. To this end, we propose a new drop-in block for video transformers - trajectory attention - that aggregates information along implicitly determined motion paths. We additionally propose a new method to address the quadratic dependence of computation and memory on the input size, which is particularly important for high resolution or long videos. While these ideas are useful in a range of settings, we apply them to the specific task of video action recognition with a transformer model and obtain state-of-the-art results on the Kinetics, Something-Something V2, and Epic-Kitchens datasets.

The results of Table 2 clearly show that the two approaches are different.

Transformers are slow to train on videos due to extremely large numbers of input tokens, even though many video tokens are repeated over time. Existing methods to remove such uninformative tokens either have significant overhead, negating any speedup, or require tuning for different datasets and examples. We present Run-Length Tokenization (RLT), a simple approach to speed up video transformers inspired by run-length encoding for data compression. RLT efficiently finds and removes'runs' of patches that are repeated over time prior to model inference, then replaces them with a single patch and a positional encoding to represent the resulting token's new length. Our method is content-aware, requiring no tuning for different datasets, and fast, incurring negligible overhead. RLT yields a large speedup in training, reducing the wall-clock time to fine-tune a video transformer by 30% while matching baseline model performance. RLT also works without any training, increasing model throughput by 35% with only 0.1% drop in accuracy. RLT speeds up training at 30 FPS by more than 100%, and on longer video datasets, can reduce the token count by up to 80%. Our project page is at https://rccchoudhury.github.io/projects/rlt/.

In video transformers, the time dimension is often treated in the same way as the two spatial dimensions. However, in a scene where objects or the camera may move, a physical point imaged at one location in frame t may be entirely unrelated to what is found at that location in frame t k . These temporal correspondences should be modeled to facilitate learning about dynamic scenes. To this end, we propose a new drop-in block for video transformers - trajectory attention - that aggregates information along implicitly determined motion paths. We additionally propose a new method to address the quadratic dependence of computation and memory on the input size, which is particularly important for high resolution or long videos. While these ideas are useful in a range of settings, we apply them to the specific task of video action recognition with a transformer model and obtain state-of-the-art results on the Kinetics, Something-Something V2, and Epic-Kitchens datasets.

We extend multimodal transformers to include 3D camera motion as a conditioning signal for the task of video generation. Generative video models are becoming increasingly powerful, thus focusing research efforts on methods of controlling the output of such models. We propose to add virtual 3D camera controls to generative video methods by conditioning generated video on an encoding of three-dimensional camera movement over the course of the generated video. Results demonstrate that we are (1) able to successfully control the camera during video generation, starting from a single frame and a camera signal, and (2) we demonstrate the accuracy of the generated 3D camera paths using traditional computer vision methods.

This paper studies the problem of concept-based interpretability of transformer representations for videos. Concretely, we seek to explain the decision-making process of video transformers based on high-level, spatiotemporal concepts that are automatically discovered. Prior research on concept-based interpretability has concentrated solely on image-level tasks. Comparatively, video models deal with the added temporal dimension, increasing complexity and posing challenges in identifying dynamic concepts over time. In this work, we systematically address these challenges by introducing the first Video Transformer Concept Discovery (VTCD) algorithm. To this end, we propose an efficient approach for unsupervised identification of units of video transformer representations - concepts, and ranking their importance to the output of a model. The resulting concepts are highly interpretable, revealing spatio-temporal reasoning mechanisms and object-centric representations in unstructured video models. Performing this analysis jointly over a diverse set of supervised and self-supervised representations, we discover that some of these mechanism are universal in video transformers. Finally, we demonstrate that VTCDcan be used to improve model performance for fine-grained tasks.

Audio and video are two most common modalities in the mainstream media platforms, e.g., YouTube. To learn from multimodal videos effectively, in this work, we propose a novel audio-video recognition approach termed audio video Transformer, AVT, leveraging the effective spatio-temporal representation by the video Transformer to improve action recognition accuracy. For multimodal fusion, simply concatenating multimodal tokens in a cross-modal Transformer requires large computational and memory resources, instead we reduce the cross-modality complexity through an audio-video bottleneck Transformer. To improve the learning efficiency of multimodal Transformer, we integrate self-supervised objectives, i.e., audio-video contrastive learning, audio-video matching, and masked audio and video learning, into AVT training, which maps diverse audio and video representations into a common multimodal representation space. We further propose a masked audio segment loss to learn semantic audio activities in AVT. Extensive experiments and ablation studies on three public datasets and two in-house datasets consistently demonstrate the effectiveness of the proposed AVT. Specifically, AVT outperforms its previous state-of-the-art counterparts on Kinetics-Sounds by 8%. AVT also surpasses one of the previous state-of-the-art video Transformers [25] by 10% on VGGSound by leveraging the audio signal. Compared to one of the previous state-of-the-art multimodal methods, MBT [32], AVT is 1.3% more efficient in terms of FLOPs and improves the accuracy by 3.8% on Epic-Kitchens-100.

As transformer architectures and dataset sizes continue to scale, the need to understand the specific dataset factors affecting model performance becomes increasingly urgent. This paper investigates how object physics attributes (color, friction coefficient, shape) and background characteristics (static, dynamic, background complexity) influence the performance of Video Transformers in trajectory prediction tasks under occlusion. Beyond mere occlusion challenges, this study aims to investigate three questions: How do object physics attributes and background characteristics influence the model performance? What kinds of attributes are most influential to the model generalization? Is there a data saturation point for large transformer model performance within a single task? To facilitate this research, we present OccluManip, a real-world video-based robot pushing dataset comprising 460,000 consistent recordings of objects with different physics and varying backgrounds. 1.4 TB and in total 1278 hours of high-quality videos of flexible temporal length along with target object trajectories are collected, accommodating tasks with different temporal requirements. Additionally, we propose Video Occlusion Transformer (VOT), a generic video-transformer-based network achieving an average 96% accuracy across all 18 sub-datasets provided in OccluManip. OccluManip and VOT will be released at: https://github.com/ShutongJIN/OccluManip.git

Transformers have become one of the dominant architectures in the field of computer vision. However, there are yet several challenges when applying such architectures to video data. Most notably, these models struggle to model the temporal patterns of video data effectively. Directly targeting this issue, we introduce Patch-Blender, a learnable blending function that operates over patch embeddings across the temporal dimension of the latent space. We show that our method is successful at enabling vision transformers to encode the temporal component of video data. On Something-Something v2 and MOVi-A, we show that our method improves the baseline performance of video Transformers. PatchBlender has the advantage of being compatible with almost any Transformer architecture and since it is learnable, the model can adaptively turn on or off the prior. It is also extremely lightweight compute-wise, 0.005% the GFLOPs of a ViT-B. The Transformer (Vaswani et al., 2017) has become one of the dominant architectures of many fields in machine learning (Brown et al., 2020; Devlin et al., 2019; Dosovitskiy et al., 2020). Initially proposed for natural language processing (Vaswani et al., 2017), it has since been shown to outperform convolutional neural networks in the image domain (Dosovitskiy et al., 2020). Adapting such vision models to the video domain has been straightforward and resulted in new state-of-the-art results (Arnab et al., 2021). Since then, multiple Transformer based methods have been proposed (Bertasius et al., 2021; Fan et al., 2021; Liu et al., 2021), making steady progress on a variety of challenges in the video domain.