Research map / Detection and segmentation / Open-set and few-shot detection
Direct Segmentation without Logits Optimization for Training-Free Open-Vocabulary Semantic Segmentation
Jiahao Li, Yang Lu, Yachao Zhang, Fangyong Wang, Yuan Xie, Yanyun Qu
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 KL divergence map between raw logits and degenerate distribution (ablation) | raw CLIP logits baseline · 4-benchmark average | mIoU: 8.9 -> 17.8 | Improved |
| add non-maximum suppression on raw logits (ablation) | KL discrepancy pipeline · 4-benchmark average | mIoU: 17.8 -> 19.1 | Improved |
| replace bilinear upsampling with joint bilateral upsampling (ablation) | KL discrepancy pipeline · 4-benchmark average | mIoU: 19.1 -> 19.6 | Improved |
| replace KL map with optimal path map (ablation) | KL + NMS + JBU pipeline · 4-benchmark average | mIoU: 19.6 -> 40.0 | Improved |
| replace KL map with maximum velocity map (ablation) | KL + NMS + JBU pipeline · 4-benchmark average | mIoU: 19.6 -> 40.8 | Improved |
| fuse optimal path and maximum velocity maps (ablation) | maximum velocity mode · 4-benchmark average | mIoU: 40.8 -> 38.4 | Hurt |
| use maximum velocity mode instead of optimal path mode | Ours · OVSS 8-benchmark average (CLIP ViT-B/16) | mIoU: 46.2 -> 46.9 | Improved |
| combine SD2 up0 and up1 attention blocks instead of up1 alone (ablation) | maximum velocity mode · 4-benchmark average | mIoU: 39.7 -> 40.8 | Improved |
| use key-key mean attention for logits instead of original attention (ablation) | CLIP logits · 4-benchmark average | mIoU: 36.2 -> 40.8 | Improved |
| use key-key instead of query-query attention for logits (ablation) | CLIP logits · 4-benchmark average | mIoU: 39.1 -> 40.8 | Improved |
| raise threshold tau above 0.3 (ablation) | maximum velocity mode · five benchmarks | mIoU: peak at tau=0.3 | Hurt |
| use more SD2 denoising steps (ablation) | maximum velocity mode · five benchmarks | mIoU | Hurt |
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 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
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.93)
- Training-free Open-Vocabulary Semantic Segmentation via Diverse Prototype Construction and Sub-region Matching (AAAI 2025 · similarity 0.92)
- Distilling Spectral Graph for Object-Context Aware Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.91)
- Transferable and Principled Efficiency for Open-Vocabulary Segmentation (CVPR 2024 · similarity 0.91)
- Open-Vocabulary Segmentation with Semantic-Assisted Calibration (CVPR 2024 · similarity 0.91)
- Relationship Prompt Learning is Enough for Open-Vocabulary Semantic Segmentation (NeurIPS 2024 · similarity 0.90)
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.
