Research map / Video / Video-language understanding
LENS: Adaptive Spatio-Temporal Zooming for Keyframe Sampling in Long-Form Videos
Ce Zhang, Jinxi He, Yaqi Xie, Katia Sycara
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 CLIP query-aware frame scoring (ablation) | Qwen2.5-VL · Video-MME overall, 8 frames | accuracy (%): 53.3 -> 56.6 | Improved |
| use BLIP instead of CLIP for frame scoring (ablation) | Qwen2.5-VL · Video-MME overall, 8 frames | accuracy (%): 56.6 -> 57.0 | Improved |
| add spatial zoom-in visual prompting (ablation) | Qwen2.5-VL · Video-MME overall, 8 frames | accuracy (%): 57.0 -> 59.6 | Improved |
| add temporal zoom-out via video graph and hyperframes (all frames) (ablation) | Qwen2.5-VL · Video-MME overall, 8 frames | accuracy (%): 57.0 -> 59.2 | Improved |
| combine spatial and temporal branches with fixed 4/4 split (ablation) | Qwen2.5-VL · Video-MME overall, 8 frames | accuracy (%): 59.6 -> 60.2 | Improved |
| add adaptive budget allocation instead of fixed 4/4 split (ablation) | Qwen2.5-VL · Video-MME overall, 8 frames | accuracy (%): 60.2 -> 60.7 | Improved |
| use stronger allocator (GPT-5-Mini or Gemini-2.5-Pro) instead of default (ablation) | LENS · Video-MME overall | accuracy (%): +0.4% and +0.7% | No clear effect |
| add graph propagation (alpha 0.8 instead of 0) (ablation) | LENS · Video-MME overall, 8 frames | accuracy (%): 60.2 -> 60.7 | Improved |
| use moderate propagation alpha 0.8 instead of 1 (ablation) | LENS · Video-MME overall, 8 frames | accuracy (%): 59.0 -> 60.7 | Improved |
| use hyperframe size tau 4 instead of 1 (ablation) | LENS · Video-MME overall, 8 frames | accuracy (%): 59.7 -> 60.7 | Improved |
| use hyperframe size tau 4 instead of 9 (ablation) | LENS · Video-MME overall, 8 frames | accuracy (%): 60.3 -> 60.7 | Improved |
| use sigma 0.3 instead of 0.1 (ablation) | LENS · Video-MME overall, 8 frames | accuracy (%): 60.4 -> 60.7 | Improved |
Closest papers by meaning
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.90)
- FOCUS: Efficient Keyframe Selection for Long Video Understanding (ICLR 2026 · similarity 0.90)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.89)
- Logic-in-Frames: Dynamic Keyframe Search via Visual Semantic-Logical Verification for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.87)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.86)
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.
