Research map / Video / Video-language understanding
Elysium: Exploring Object-level Perception in Videos through Semantic Integration Using MLLMs
Han Wang, Yanjie Wang, Ye Yongjie, Yuxiang Nie, Can 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 |
|---|---|---|---|
| use T-Selector token selection instead of concatenation at equal 144 tokens (ablation) | Elysium · RefCOCO series average | accuracy: 54.65 -> 78.45 | Improved |
| use T-Selector instead of one-layer cross-attention at equal 144 tokens (ablation) | Elysium · RefCOCO series average | accuracy: 49.23 -> 78.45 | Improved |
| use 336px ViT with T-Selector instead of 224px ViT with MLP at equal token count (ablation) | Elysium · RefCOCO series average | accuracy: 76.81 -> 80.09 | Improved |
| reduce visual tokens per image with T-Selector (108 to 72) (ablation) | Elysium · RefCOCO series average | accuracy: 78.54 -> 76.29 | Hurt |
| increase visual tokens from 108 to 256 with T-Selector (ablation) | Elysium · RefCOCO series average | accuracy: 78.54 -> 80.09 | Improved |
| compress to 108 tokens with T-Selector versus 576 uncompressed tokens at 336px (ablation) | Elysium · RefCOCO series average | accuracy: 81.45 -> 78.54 | Hurt |
| train end-to-end video MLLM on ElysiumTrack-1M instead of per-frame MiniGPT-v2 | Elysium · ElysiumTrack-1M eval RSOT | AUC: 65.4 -> 87.5 | Improved |
| zero-shot tracking on datasets with small objects | Elysium · UAV123 | AUC: 56.6 vs 64.6 (SiamGAT) | Hurt |
Later papers that compared with this method
1 later accepted paper reports Elysium as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- Omni-RGPT: Unifying Image and Video Region-level Understanding via Token Marks (CVPR 2025 · better on 0 of 1 shared result)
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.
- MovieChat — MovieChat: From Dense Token to Sparse Memory for Long Video Understanding (CVPR 2024) · better on 2 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
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.90)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.89)
- Same or Not? Enhancing Visual Perception in Vision-Language Models (CVPR 2026 · similarity 0.89)
- Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis (CVPR 2025 · similarity 0.88)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.88)
- MVBench: A Comprehensive Multi-modal Video Understanding Benchmark (CVPR 2024 · 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.
