Research map / Video / Video-language understanding
ReMoRa: Multimodal Large Language Model based on Refined Motion Representation for Long-Video Understanding
Daichi Yashima, Shuhei Kurita, Yusuke Oda, Komei Sugiura
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 |
|---|---|---|---|
| use codec-aware 64 I-frames instead of uniform 64 frames (ablation) | ReMoRa · VideoMME | accuracy: 62.4 -> 64.3 | Improved |
| increase number of I-frames from 16 to 64 (ablation) | ReMoRa · VideoMME | accuracy: 58.5 -> 64.3 | Improved |
| increase number of I-frames from 16 to 64 (ablation) | ReMoRa · NExT-QA | accuracy: 81.6 -> 84.2 | Improved |
| add optical-flow pretraining of RMR (ablation) | ReMoRa · NExT-QA | accuracy: 82.2 -> 84.2 | Improved |
| add RMR module (ablation) | ReMoRa · VideoMME | accuracy: 62.1 -> 64.3 | Improved |
| use HMSS instead of cross-attention fusion (ablation) | ReMoRa · VideoMME | accuracy: 62.5 -> 64.3 | Improved |
| use HMSS instead of additive fusion (ablation) | ReMoRa · NExT-QA | accuracy: 81.5 -> 84.2 | Improved |
| use ReMoRa instead of LLaVA-Video for memory use | video MLLM inference · NExT-QA 50 videos (H200) | peak GPU memory (GB): 23.21 -> 10.59 | Improved |
| use ReMoRa instead of LLaVA-Video for throughput | video MLLM inference · NExT-QA 50 videos (H200) | samples/s: 0.53 -> 0.40 | Hurt |
| re-encode with harsh H.264 settings (CRF 35, GOP 64, no B-frames, 256p) (ablation) | ReMoRa · NExT-QA | accuracy: 84.9 -> 81.5 | Hurt |
| vary GOP size or B-frames in H.264 encoding (ablation) | ReMoRa · NExT-QA | accuracy: 84.9 -> 84.3 to 84.6 | No clear effect |
Closest papers by meaning
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.90)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.90)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.89)
- HIMMEL: Hierarchical Interleaved Multi-stream Motion Encoding for Long Video Understanding (NeurIPS 2026 · similarity 0.89)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.88)
- Understanding Long Videos with Multimodal Language Models (ICLR 2025 · similarity 0.88)
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.
