Research map / Detection and segmentation / Open-set and few-shot detection
The Power of Prior: Training-Free Open-Vocabulary Semantic Segmentation with LLaVA
Bingfeng Zhang, Siyue Yu, Hui Li, Jiahua Lin, Wenwu Wang, Jimin Xiao
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 |
|---|---|---|---|
| use query 'Does the image have the class of {}?' vs 'Does the image contain a {}?' (ablation) | FSeg-LLaVA1.6 · VOC21 | mIoU: 62.4 -> 65.9 | Improved |
| use visual responses from LLM layers 7-13 instead of 15-22 (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 49.1 -> 68.0 | Improved |
| use layers 7-13 instead of deep layers 22-30 (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 15.2 -> 68.0 | Improved |
| use layers 7-13 instead of 7-17 (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 66.8 -> 68.0 | Improved |
| use both point and box prompts for SAM instead of points only (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 61.4 -> 68.0 | Improved |
| use point prompts instead of box prompts (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 57.8 -> 61.4 | Improved |
| filter distance map D with attention map MA (Df = D·MA) instead of D alone (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 67.3 -> 68.0 | Improved |
| use attention map alone instead of D·MA (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 68.0 -> 27.3 | Hurt |
| multiply prototype mask Mv with guidance map Mf instead of Mf alone (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 67.3 -> 68.0 | Improved |
| add prototype mask Mv to refinement instead of Mv alone (ablation) | FSeg-LLaVA1.5 · VOC21 | mIoU: 48.4 -> 68.0 | Improved |
| scale LLM from Vicuna-7B to Vicuna-13B in LLaVA-1.6 | FSeg-LLaVA1.6 · VOC21 | mIoU: 65.9 -> 64.1 | Hurt |
| upgrade from LLaVA-1.5 to LLaVA-1.6 (Vicuna-7B) | FSeg-LLaVA · Average over 5 benchmarks | mIoU: 35.7 -> 36.8 | Mixed |
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.
- ProxyCLIP — ProxyCLIP: Proxy Attention Improves CLIP for Open-Vocabulary Segmentation (ECCV 2024) · better on 3 of 5 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
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.91)
- OpenWorldSAM: Extending SAM2 for Universal Image Segmentation with Language Prompts (NeurIPS 2025 · similarity 0.89)
- LRM-LLaVA: Overcoming the Modality Gap of Multilingual Large Language-Vision Model for Low-Resource Languages (AAAI 2025 · similarity 0.89)
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation (NeurIPS 2025 · similarity 0.89)
- DSV-LFS: Unifying LLM-Driven Semantic Cues with Visual Features for Robust Few-Shot Segmentation (CVPR 2025 · similarity 0.89)
- LLaFS: When Large Language Models Meet Few-Shot Segmentation (CVPR 2024 · 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.
