Research map / Video / Video-language understanding
ViLLa: Video Reasoning Segmentation with Large Language Model
Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Hengshuang Zhao
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 the context synthesizer (ablation) | ViLLa baseline · VideoReasonSeg | accuracy: 39.7 -> 43.3 | Improved |
| add the key segment extractor (ablation) | ViLLa baseline · VideoReasonSeg | accuracy: 39.7 -> 43.8 | Improved |
| add the hierarchical temporal synchronizer (ablation) | ViLLa baseline · VideoReasonSeg | J&F: 48.5 -> 50.7 | Improved |
| combine all three modules (ablation) | ViLLa baseline · VideoReasonSeg | J&F: 48.5 -> 55.4 | Improved |
| add multi-scale segmentation tokens only (without synchronization) (ablation) | ViLLa · VideoReasonSeg | J&F: 53.7 -> 54.2 | No clear effect |
| add video-frame synchronization on top of multi-scale tokens (ablation) | ViLLa · VideoReasonSeg | J&F: 54.2 -> 55.4 | Improved |
| use 4 aggregation layers rather than 3 (ablation) | ViLLa · VideoReasonSeg | J&F: 55.4 -> 55.1 | Hurt |
| use InternVideo2 as vision backbone instead of SAM (ablation) | ViLLa · VideoReasonSeg | J&F: 50.5 -> 55.4 | Improved |
| adaptive sampling (1/3 global, 2/3 neighbor) instead of global or neighbor only (ablation) | Key segment extractor · VideoReasonSeg | J&F: 54.7 -> 55.4 | Improved |
| use embedding-similarity aggregation instead of PixelLM-style feature fusion (ablation) | ViLLa · Ref-DAVIS17 | J&F: 63.2 -> 64.4 | Improved |
| use ViLLa dedicated design instead of direct adaptation of LISA-style modules (ViLLa dagger) (ablation) | ViLLa · VideoReasonSeg | J&F: 51.9 -> 55.4 | 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.
- VISA — VISA: Reasoning Video Object Segmentation via Large Language Model (ECCV 2024) · better on 3 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
- The Devil is in Temporal Token: High Quality Video Reasoning Segmentation (CVPR 2025 · similarity 0.91)
- One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos (NeurIPS 2024 · similarity 0.91)
- Fast Reasoning Segmentation for Images and Videos (CVPR 2026 · similarity 0.89)
- LISA: Reasoning Segmentation via Large Language Model (CVPR 2024 · similarity 0.89)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.89)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.88)
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.
