Research map / Video / Video-language understanding
VideoITG: Multimodal Video Understanding with Instructed Temporal Grounding
Shihao Wang, Guo Chen, De-An Huang, Zhiqi Li, Minghan Li, Guilin Liu, Jan Kautz, Jose M. Alvarez, Lei Zhang, Zhiding Yu
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 uniform sampling with VideoITG frame selection | LLaVA-OneVision-7B · LongVideoBench/MLVU/VideoMME avg | average accuracy (%): 54.9 -> 61.6 | Improved |
| replace uniform sampling with VideoITG frame selection | Qwen2-VL · LongVideoBench/MLVU/VideoMME avg | average accuracy (%): 54.3 -> 60.3 | Improved |
| replace uniform 32 frames with VideoITG top-32 | six Video-LLMs (InternVL2.5/3.5, Qwen3-VL, LLaVA-Video, Eagle2.5) · LongVideoBench/MLVU/VideoMME/CG-Bench avg | average accuracy (%): 58.7 -> 64.3 (InternVL2.5-8B); 61.8 -> 66.7 (Eagle2.5-8B) | Improved |
| use VideoITG at 32 frames instead of uniform 64 frames | LLaVA-Video-7B · LongVideoBench/MLVU/VideoMME avg | average accuracy (%): 64.7 -> 66.9 | Improved |
| use VideoITG instead of QuoTA (both better than uniform) | LLaVA-Video-7B · LongVideoBench/MLVU/VideoMME avg | average accuracy (%): 65.9 -> 67.1 | Improved |
| use text-generation formulation instead of classification (ablation) | VideoITG · VideoMME/MLVU/LongVideoBench avg | average accuracy (%): 67.8 -> 48.5 | Hurt |
| use full bidirectional attention instead of causal+anchor (ablation) | VideoITG · VideoMME/MLVU/LongVideoBench avg | average accuracy (%): 67.2 -> 67.8 | Improved |
| scale grounding model from 3B to 7B (ablation) | VideoITG · VideoMME/MLVU/LongVideoBench avg | average accuracy (%): 66.8 -> 67.8 | Improved |
| add instructed clip captioning to annotation pipeline (ablation) | VideoITG-40K · VideoMME Long | accuracy (%): 53.4 -> 56.9 | Improved |
| add instructed frame localization to annotation pipeline (ablation) | VideoITG-40K · VideoMME Medium | accuracy (%): 65.8 -> 67.1 | Improved |
| add video data to pre-training (ablation) | VideoITG · VideoMME/MLVU/LongVideoBench avg | average accuracy (%): 67.1 -> 67.8 | No clear effect |
| add image and video alignment pre-training (ablation) | VideoITG · MLVU | accuracy (%): 69.1 -> 75.0 | 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.
- Frame-VOYAGER — Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025) · better on 1 of 1 shared result
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
- TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning (ICLR 2025 · similarity 0.90)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.90)
- VideoGLaMM : A Large Multimodal Model for Pixel-Level Visual Grounding in Videos (CVPR 2025 · similarity 0.90)
- SnAG: Scalable and Accurate Video Grounding (CVPR 2024 · similarity 0.89)
- Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding (NeurIPS 2025 · similarity 0.89)
- TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024 · similarity 0.89)
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.
