Research map / Detection and segmentation / Open-set and few-shot detection
Side Adapter Network for Open-Vocabulary Semantic Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 end-to-end CLIP-aware mask prediction instead of blocking gradients from CLIP (ablation) | SAN · ADE20K-150 | mIoU: 21.6 -> 26.1 (+4.5) | Helped |
| Decouple the mask proposal and recognition heads (ablation) | SAN · ADE20K-150 | mIoU: 26.1 -> 27.8 (+1.7) | Helped |
| Add CLIP feature fusion rather than removing it (ablation) | SAN · ADE20K-150 | mIoU: 21.1 -> 27.8 | Helped |
| Fuse deeper CLIP features rather than only the stem feature (ablation) | SAN · ADE20K-150 | mIoU: 20.0 -> 27.1 | Helped |
| Fuse features from multiple CLIP layers rather than only layer 9 (ablation) | SAN · ADE20K-150 | mIoU: +0.8 mIoU | Helped |
| Use single-forward instead of twice-forward extraction (ablation) | SAN · ADE20K-150 | mIoU: 27.6 vs 27.8; 0.2 mIoU gap | No clear effect |
| Allocate 9 layers to feature fusion and 3 to recognition (ablation) | SAN · ADE20K-150 | mIoU: 27.8 | Helped |
| Use CLIP input resolution 320² instead of 224² (ablation) | SAN · ADE20K-150 | mIoU: 26.3 -> 27.8 | Helped |
| Increase CLIP input resolution to 448² or 640² instead of 320² (ablation) | SAN · ADE20K-150 | mIoU: 27.8 -> 26.1; 27.8 -> 24.6 | Hurt |
| Fine-tune CLIP positional embeddings rather than keeping them fixed (ablation) | SAN · ADE20K-150 | mIoU: 27.0 -> 27.8 | Helped |
| Use 240-dimensional SAN rather than 144-dimensional SAN (ablation) | SAN · ADE20K-150 | mIoU: 26.7 -> 27.8 | Helped |
| Increase SAN width from 240 to 288 (ablation) | SAN · ADE20K-150 | mIoU: 27.8 -> 27.3 | Hurt |
Closest papers by meaning
- Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation (CVPR 2025 · similarity 0.91)
- Effective SAM Combination for Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.90)
- Unveiling the Knowledge of CLIP for Training-Free Open-Vocabulary Semantic Segmentation (AAAI 2025 · similarity 0.89)
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.89)
- CLIP as RNN: Segment Countless Visual Concepts without Training Endeavor (CVPR 2024 · similarity 0.88)
- Towards Open-Vocabulary Semantic Segmentation Without Semantic Labels (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.