Keeping Your Eye on the Ball: Trajectory Attention in Video Transformers
This work addresses the challenge of dynamic scene understanding in video analysis for applications like action recognition, though it is incremental as it builds on existing transformer architectures.
The paper tackles the problem of modeling temporal correspondences in video transformers by proposing trajectory attention, which aggregates information along motion paths, and addresses computational efficiency for high-resolution or long videos. The method achieves state-of-the-art results on Kinetics, Something-Something V2, and Epic-Kitchens datasets for video action recognition.
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. Code and models are available at: https://github.com/facebookresearch/Motionformer