Research map / Video / Video-language understanding
IPFormer: Instance Prompt-guided Transformer for Multi-modal Multi-shot Video Understanding
Yujia Liang, Jile Jiao, Xuetao Feng, Xinchen Liu, Kun Liu, Yuan Wang, Zixuan Ye, Hao Lu, Zhicheng Wang
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 frame queries to Q-Former learnable queries (ablation) | IPFormer compressor · ActivityNet-QA | accuracy: 40.9 -> 42.6 | Improved |
| add instance queries without clustering (ablation) | IPFormer compressor · ActivityNet-QA | accuracy: 42.6 -> 43.6 | Improved |
| cluster instance tokens (ablation) | IPFormer compressor · ActivityNet-QA | accuracy: 43.6 -> 45.5 | Improved |
| cluster instance tokens (ablation) | IPFormer compressor · MultiClip-Bench | accuracy: 44.6 -> 46.2 | Improved |
| replace full projection by IPFormer compressor | Video-LLaVA · MultiClip-Bench | accuracy: 44.5 -> 46.2 | Improved |
| use plain Q-Former instead of full projection (ablation) | Video-LLaVA · MultiClip-Bench | accuracy: 44.5 -> 42.3 | Hurt |
| use 2x2 average pooling instead of full projection (ablation) | Video-LLaVA · MultiClip-Bench | accuracy: 44.5 -> 43.8 | Hurt |
| add IPFormer model design (no MultiClip data) | Video-LLaVA · MultiClip-Bench | mean accuracy: 44.5 -> 48.1 | Improved |
| add MultiClip training data (ablation) | IPFormer · MultiClip-Bench | mean accuracy: 48.1 -> 57.0 | Improved |
| add VideoChat2 instruction data (353k) (ablation) | IPFormer · ActivityNet-QA | accuracy: 45.8 -> 48.5 | Improved |
| add MultiClip training data (ablation) | PLLaVA · MultiClip-Bench | accuracy: 49.4 -> 57.6 | Improved |
| increase frames from 16 to 48 | IPFormer · MultiClip-Bench | mean accuracy: 57.0 -> 58.8 | Improved |
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.
- ST-LLM — ST-LLM: Large Language Models Are Effective Temporal Learners (ECCV 2024) · better on 0 of 3 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
- MVBench: A Comprehensive Multi-modal Video Understanding Benchmark (CVPR 2024 · similarity 0.86)
- VideoPrism: A Foundational Visual Encoder for Video Understanding (ICML 2024 · similarity 0.86)
- Vript: A Video Is Worth Thousands of Words (NeurIPS 2024 · similarity 0.86)
- HierarQ: Task-Aware Hierarchical Q-Former for Enhanced Video Understanding (CVPR 2025 · similarity 0.86)
- SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding (CVPR 2025 · similarity 0.86)
- InstanceCap: Improving Text-to-Video Generation via Instance-aware Structured Caption (CVPR 2025 · similarity 0.85)
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.
