Research map / Video / Video-language understanding
Natural Language Camera Movement Understanding
Yuwen Tan, Joey Huang, Jin Huang, Haoxiang Li, Boqing Gong
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 targeted re-sampling to balance categories (ablation) | SFT Qwen3-VL-4B · ACaM real-world | average class accuracy: 48.42 -> 63.78 | Improved |
| add targeted re-sampling to balance categories (ablation) | SFT Qwen3-VL-4B · ACaM synthetic | average class accuracy: 58.56 -> 72.07 | Improved |
| add data augmentation (zoom crop, reversal, flip, roll) on top of sampling (ablation) | SFT Qwen3-VL-4B · ACaM real-world | average class accuracy: 63.78 -> 70.88 | Improved |
| add data augmentation on top of sampling (ablation) | SFT Qwen3-VL-4B · ACaM synthetic | average class accuracy: 72.07 -> 75.03 | Improved |
| raise LoRA rank from 64 to 256 (ablation) | SFT Qwen3-VL-4B · ACaM real-world | average class accuracy: 70.88 -> 72.27 | Improved |
| raise LoRA rank from 64 to 256 (ablation) | SFT Qwen3-VL-4B · ACaM synthetic | average class accuracy: 75.03 -> 74.40 | No clear effect |
| raise LoRA rank from 64 to 256 (ablation) | SFT Qwen3-VL-8B · ACaM real-world | overall accuracy: 68.10 -> 72.75 | Improved |
| raise LoRA rank from 64 to 256 (ablation) | SFT Qwen3-VL-8B · ACaM synthetic | overall accuracy: 70.91 -> 78.20 | Improved |
| scale SFT model from 4B to 8B | SFT Qwen3-VL · ACaM synthetic | overall accuracy: 73.28 -> 78.20 | Improved |
| instruction-tune Qwen3-VL-8B on ACaM training set | Qwen3-VL-8B · ACaM real-world | overall accuracy: 58.27 -> 72.75 | Improved |
| increase input frames from 8 to 32 for low-intensity push in (ablation) | Gemini-3-Pro · Push-in low intensity | accuracy: 29.17 -> 29.17 | No clear effect |
| increase camera movement intensity from low to high | Qwen3-VL-32B · Push-in (8 frames) | accuracy: 41.67 -> 100.00 | 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.
- ShotVL-7B — ShotVL: Human-Centric Highlight Frame Retrieval via Language Queries (AAAI 2025) · better on 2 of 2 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
- Towards Understanding Camera Motions in Any Video (NeurIPS 2025 · similarity 0.90)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.88)
- PixFoundation 2.0: Do Video Multi-Modal LLMs Use Motion in Visual Grounding? (NeurIPS 2026 · similarity 0.87)
- FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding (NeurIPS 2025 · similarity 0.87)
- Understanding Long Videos with Multimodal Language Models (ICLR 2025 · similarity 0.87)
- ShotBench: Expert-Level Cinematic Understanding in Vision-Language Models (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.
