English

Learning Trajectory-Word Alignments for Video-Language Tasks

Computer Vision and Pattern Recognition 2023-03-10 v3

Abstract

In a video, an object usually appears as the trajectory, i.e., it spans over a few spatial but longer temporal patches, that contains abundant spatiotemporal contexts. However, modern Video-Language BERTs (VDL-BERTs) neglect this trajectory characteristic that they usually follow image-language BERTs (IL-BERTs) to deploy the patch-to-word (P2W) attention that may over-exploit trivial spatial contexts and neglect significant temporal contexts. To amend this, we propose a novel TW-BERT to learn Trajectory-Word alignment by a newly designed trajectory-to-word (T2W) attention for solving video-language tasks. Moreover, previous VDL-BERTs usually uniformly sample a few frames into the model while different trajectories have diverse graininess, i.e., some trajectories span longer frames and some span shorter, and using a few frames will lose certain useful temporal contexts. However, simply sampling more frames will also make pre-training infeasible due to the largely increased training burdens. To alleviate the problem, during the fine-tuning stage, we insert a novel Hierarchical Frame-Selector (HFS) module into the video encoder. HFS gradually selects the suitable frames conditioned on the text context for the later cross-modal encoder to learn better trajectory-word alignments. By the proposed T2W attention and HFS, our TW-BERT achieves SOTA performances on text-to-video retrieval tasks, and comparable performances on video question-answering tasks with some VDL-BERTs trained on much more data. The code will be available in the supplementary material.

Keywords

Cite

@article{arxiv.2301.01953,
  title  = {Learning Trajectory-Word Alignments for Video-Language Tasks},
  author = {Xu Yang and Zhangzikang Li and Haiyang Xu and Hanwang Zhang and Qinghao Ye and Chenliang Li and Ming Yan and Yu Zhang and Fei Huang and Songfang Huang},
  journal= {arXiv preprint arXiv:2301.01953},
  year   = {2023}
}
R2 v1 2026-06-28T08:03:27.392Z