Research map / Detection and segmentation / Open-set and few-shot detection
Exploring Open-Vocabulary Semantic Segmentation from CLIP Vision Encoder Distillation Only
Jun Chen, Deyao Zhu, Guocheng Qian, Bernard Ghanem, Zhicheng Yan, Chenchen Zhu, Fanyi Xiao, Sean Chang Culatana, Mohamed Elhoseiny
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 multi-scale (2x2,3x3,4x4) CLIP feature distillation (ablation) | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 21.1 -> 28.5 | Improved |
| add segment matching loss (ablation) | ZeroSeg base · PASCAL VOC 2012 (zero-shot) | mIoU: 21.1 -> 38.6 | Improved |
| add segment matching loss to multi-scale model (ablation) | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 28.5 -> 40.8 | Improved |
| use all grid scales 1x1+2x2+3x3+4x4 instead of 1x1+3x3 (ablation) | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 40.2 -> 40.8 | Improved |
| use 60% encoder mask ratio instead of 0% (ablation) | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 35.6 -> 40.8 | Improved |
| raise mask ratio from 60% to 70% (ablation) | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 40.8 -> 33.3 | Hurt |
| train on ImageNet-1K instead of CC3M+COCO | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 37.3 -> 40.8 | Improved |
| scale training data to CC12M+IN-1K | ZeroSeg · PASCAL VOC 2012 (zero-shot) | mIoU: 40.8 -> 42.9 | Improved |
| replace text-supervised GroupViT with CLIP-vision-distilled ZeroSeg | open-vocabulary segmentation · Human study, 1000-class vocabulary | votes: 323/1000 -> 677/1000 | Improved |
| replace GroupViT with ZeroSeg for compute | training cost · Training cost (V100) | GPU hours: ~768 -> ~84 | Improved |
| ZeroSeg (IN-1K) versus GroupViT (CC12M+YFCC) | PASCAL VOC / Context / COCO · COCO (zero-shot) | mIoU: 24.3 -> 20.2 | 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.
- GroupViT* — GroupViT: Semantic Segmentation Emerges From Text Supervision (CVPR 2022) · better on 4 of 4 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
- A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-Language Model (ECCV 2022 · similarity 0.92)
- Open-World Semantic Segmentation via Contrasting and Clustering Vision-Language Embedding (ECCV 2022 · similarity 0.92)
- Exploring Regional Clues in CLIP for Zero-Shot Semantic Segmentation (CVPR 2024 · similarity 0.91)
- Exploring Simple Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.91)
- A Simple Framework for Open-Vocabulary Zero-Shot Segmentation (ICLR 2025 · similarity 0.91)
- Cascade-CLIP: Cascaded Vision-Language Embeddings Alignment for Zero-Shot Semantic Segmentation (ICML 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.
