Research map / Video / Video-language understanding
Video Panels for Long Video Understanding
Lars Doorenbos, Federico Spurio, Juergen Gall
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 |
|---|---|---|---|
| apply 2x2 video panels to LLaVA-OV 7B | LLaVA-OV 7B · 5-benchmark average | accuracy: 52.8 -> 56.2 | Improved |
| apply panels to VideoLLaMA 3 7B | VideoLLaMA 3 7B · TimeScope Long | accuracy: 39.1 -> 46.7 | Improved |
| apply panels to VideoLLaMA 3 7B | VideoLLaMA 3 7B · MF2 | accuracy: -0.6 | Hurt |
| apply panels to long-context Qwen-2.5VL 7B | Qwen-2.5VL 7B (180 frames) · 5-benchmark average | accuracy: 59.7 -> 60.5 | Improved |
| replace token pooling with frame panels (ablation) | LLaVA-OV 7B · TimeScope | accuracy: 68.7 -> 69.5 | Improved |
| replace token pooling with frame panels (ablation) | LLaVA-Video 7B · TimeScope | accuracy: 79.2 vs 79.2 (low-res 78.4) | No clear effect |
| fine-tune projector+LLM with panels (vs zero-shot panels) (ablation) | LLaVA-OV 7B · VideoMME overall | accuracy: 58.9 -> 59.3 | Improved |
| fine-tune with standard inputs instead of panels (ablation) | LLaVA-OV 7B · VideoMME overall | accuracy: 58.5 -> 58.5 | No clear effect |
| use 4x4 panels instead of 2x2 (ablation) | LLaVA-OV 7B · TimeScope Long | accuracy: 33.8 -> 30.9 | Hurt |
| use uneven panels (alpha != beta) (ablation) | LLaVA-OV 7B · TimeScope Short | accuracy: 69.5 -> 65.9 (1x2), 63.5 (2x1) | Hurt |
| paneling threshold gamma 3xfps instead of 1xfps (ablation) | LLaVA-OV 7B · TimeScope Short | accuracy: 69.5 -> 67.7 | Hurt |
| apply panels to GPT-4o-mini and GPT-4.1 | GPT models · VideoMME overall | accuracy: 51.1 -> 53.6; 72.5 -> 73.6 | Improved |
Closest papers by meaning
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.92)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.92)
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.91)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.91)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.91)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.90)
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.
