Research map / Video / Video-language understanding
Learning Trajectory-Word Alignments for Video-Language Tasks
Xu Yang, Zhangzikang Li, Haiyang Xu, Hanwang Zhang, Qinghao Ye, Chenliang Li, Ming Yan, Yu Zhang, Fei Huang, Songfang Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace patch-to-word attention with trajectory-to-word (T2W) attention (ablation) | TW-BERT cross-modal encoder · MSRVTT zero-shot | R@5: 46.4 -> 50.0 | Improved |
| replace P2W with T2W attention (ablation) | TW-BERT · DiDeMo zero-shot | R@1: 26.6 -> 28.4 | Improved |
| replace P2W with T2W attention (ablation) | TW-BERT · MSVD-QA | accuracy: 47.4 -> 48.2 | Improved |
| use T2W trajectory attention instead of mean-pooling trajectory over time (ablation) | TW-BERT · DiDeMo zero-shot | R@1: 27.4 -> 28.4 | Improved |
| place frame-selector layers deeper ([6,12] instead of [2,4]) (ablation) | HFS · MSRVTT fine-tuning | R@1: 36.7 -> 38.4 | Improved |
| use HFS (F@20-14-8) rather than uniform 8 frames (ablation) | TW-BERT fine-tuning · MSRVTT fine-tuning | R@1: 38.1 -> 38.4 | Improved |
| use HFS (F@32-24-16) rather than uniform 16 frames (ablation) | TW-BERT fine-tuning · MSVD-QA | accuracy: 48.2 -> 48.5 | Improved |
| increase sampled frames from 8-6-4 to 20-14-8 with HFS (ablation) | TW-BERT · MSRVTT fine-tuning | R@1: 37.5 -> 38.4 | Improved |
| TW-BERT with 5.5M pretrain pairs versus VIOLET with 186M | TW-BERT · MSRVTT-QA | accuracy: 43.9 -> 43.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MILES — MILES: Visual BERT Pre-training with Injected Language Semantics for Video-Text Retrieval (ECCV 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Video-Panda: Parameter-efficient Alignment for Encoder-free Video-Language Models (CVPR 2025 · similarity 0.86)
- VidLA: Video-Language Alignment at Scale (CVPR 2024 · similarity 0.85)
- Kronecker Mask and Interpretive Prompts are Language-Action Video Learners (ICLR 2025 · similarity 0.85)
- TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment (NeurIPS 2024 · similarity 0.85)
- Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding (NeurIPS 2025 · similarity 0.85)
- Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
