Research map / Video / Video-language understanding
DynImg: Key Frames with Visual Prompts are Good Representation for Multi-Modal Video Understanding
Xiaoyi Bao, Chenwei Xie, Hao Tang, Tingyu Weng, Xiaofeng Wang, Yun Zheng, Xingang Wang
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 |
|---|---|---|---|
| add temporal prompts before the visual encoder (ablation) | DynImg · MSVD-QA | accuracy: 74.9 -> 77.3 | Improved |
| apply temporal prompts before the encoder instead of after (ablation) | DynImg · MSVD-QA | accuracy: 75.2 -> 77.3 | Improved |
| add temporal prompts after the encoder (ablation) | DynImg · MSVD-QA | accuracy: 74.9 -> 75.2 | No clear effect |
| add 4D RoPE (ablation) | DynImg · MSVD-QA | accuracy: 77.3 -> 78.6 | Improved |
| use 4 N-frames instead of 1 N-frame (ablation) | DynImg · MSVD-QA | accuracy: 71.9 -> 78.6 | Improved |
| increase N-frames from 4 to 12 (ablation) | DynImg · MSVD-QA | accuracy: 78.6 -> 77.5 | Hurt |
| use 4 DynImgs instead of 1 (ablation) | DynImg · MSVD-QA | accuracy: 77.0 -> 78.6 | Improved |
| increase DynImgs from 4 to 16 (ablation) | DynImg · MSVD-QA | accuracy: 78.6 -> 77.7 | Hurt |
| use 4D coordinates instead of 1D coordinates (ablation) | DynImg · MSVD-QA | accuracy: 77.3 -> 78.6 | Improved |
| make theta trainable instead of fixed in 4D split RoPE (ablation) | DynImg · MSVD-QA | accuracy: 77.0 -> 78.4 | Improved |
| merge angles instead of splitting trigonometric terms (ablation) | DynImg · MSVD-QA | accuracy: 78.4 -> 78.6 | Improved |
| DynImg with 4 DynImgs versus PLLaVA with 16 frames | DynImg · average of four open-ended QA sets | accuracy: comparable accuracy with 4 frames vs 16 frames | Improved |
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.
- ST-LLM — ST-LLM: Large Language Models Are Effective Temporal Learners (ECCV 2024) · better on 5 of 5 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
- STOP: Integrated Spatial-Temporal Dynamic Prompting for Video Understanding (CVPR 2025 · similarity 0.89)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.89)
- VideoRefer Suite: Advancing Spatial-Temporal Object Understanding with Video LLM (CVPR 2025 · similarity 0.88)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.88)
- LLaVA-ST: A Multimodal Large Language Model for Fine-Grained Spatial-Temporal Understanding (CVPR 2025 · similarity 0.88)
- State Space Prompting via Gathering and Spreading Spatio-Temporal Information for Video Understanding (NeurIPS 2025 · similarity 0.87)
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.
