Research map / Video / Video-language understanding
FrameOracle: Learning What to See and How Much to See in Videos
Chaoyu Li, Tianzhi Li, Fei Tao, ZHENYU ZHAO, Ziqian Wu, Maozheng Zhao, Juntong Song, Cheng Niu, Pooyan Fazli
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 FrameOracle selection (16 -> 10.4 frames) | LLaVA-Video-7B · 6 video benchmarks avg | avg accuracy: 54.6 -> 54.7 | No clear effect |
| add FrameOracle selection from 64 candidates (13.9 frames) | LLaVA-OneVision-7B · 6 video benchmarks avg | avg accuracy: 49.8 -> 51.6 | Improved |
| select frames with FrameOracle instead of uniform 10 of 16 (ablation) | LLaVA-OneVision-7B · avg accuracy | accuracy: 46.3 -> 50.3 | Improved |
| use FrameOracle selection on 16-frame input | LLaVA-OneVision-7B · NExTQA OE val | accuracy: 14.6 -> 16.1 | Improved |
| add FrameOracle (16 -> 10.4 frames) | LLaVA-Video-7B · cost per sample | latency: 0.615 -> 0.363 s | Improved |
| increase raw frame budget from 16 to 64 | LLaVA-Video-7B · avg accuracy | avg accuracy: 54.6 -> 56.6 | Improved |
| add Stage 1 text-visual alignment alone (ablation) | FrameOracle · 6 benchmarks avg | avg accuracy: 50.5 -> 48.5 | Hurt |
| add Stage 2 frame ranking (ablation) | FrameOracle · Qwen2.5-VL-3B | avg accuracy: 48.5 -> 50.8 | Improved |
| add Stage 3 K Head calibration (ablation) | FrameOracle · Qwen2.5-VL-3B | avg accuracy: 46.9 -> 52.8 | Improved |
| add Stage 4 fine-tuning on FrameOracle-41K (ablation) | FrameOracle · Qwen2.5-VL-3B avg | avg accuracy: 52.7 -> 52.8 | Improved |
| train Stage 2 and 3 jointly instead of staged (ablation) | FrameOracle · Rank head quality | Kendall tau: 0.5367 -> 0.2313 | Hurt |
| skip intermediate stages (Stage 1+4 only, fixed K) (ablation) | FrameOracle · Qwen2.5-VL-3B avg | avg accuracy: 50.8 -> 49.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- + Frame-Voyager — Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025) · better on 0 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
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.90)
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.90)
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.89)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.89)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.89)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · 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.
