Research map / Detection and segmentation / Open-set and few-shot detection
Scaling Open-Vocabulary Image Segmentation with Image-Level Labels
Golnaz Ghiasi, Xiuye Gu, Yin Cui, Tsung-Yi Lin
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 predicted mask proposals at inference (ablation) | OpenSeg (COCO) · PASCAL-Context-59 / ADE20k-150 | mIoU: 32.1 -> 42.1 (PC-59); 16.4 -> 21.1 (A-150) | Improved |
| use ground-truth masks as proposals instead of predicted masks | OpenSeg (COCO) · PASCAL-Context-59 / ADE20k-150 | mIoU: 42.1 -> 49.3 (PC-59); 21.1 -> 27.5 (A-150) | Improved |
| add class-agnostic segmentation loss to grounding loss (ablation) | OpenSeg (COCO) · PASCAL-Context-59 / ADE20k-847 | mIoU: 15.6 -> 42.1 (PC-59); 2.5 -> 6.3 (A-847) | Improved |
| scale training data from COCO to COCO+Localized Narratives with pseudo mask labels | OpenSeg · ADE20k-150 / PASCAL-Context-459 | mIoU: 21.1 -> 24.8 (A-150); 9.0 -> 11.2 (PC-459) | Improved |
| compute segmentation loss on pseudo-labeled Localized Narratives examples (ablation) | OpenSeg (COCO+Loc. Narr.) · PASCAL-Context-59 / ADE20k-150 | mIoU: 47.3 -> 45.9 (PC-59); 22.4 -> 24.8 (A-150) | Mixed |
| add ensembling and prompt engineering of text queries | OpenSeg (EfficientNet-B7) · ADE20k-150 | mIoU: 26.4 -> 28.6 | Improved |
| replace per-pixel class-label training (LSeg+) with caption supervision (OpenSeg), ResNet101 COCO | LSeg+ · ADE20k-150 | mIoU: 13.0 -> 15.3 | Improved |
| fine-tune ALIGN with per-pixel COCO labels (LSeg+) | ALIGN · ADE20k-847 | mIoU: 1.0 mIoU worse than ALIGN | Hurt |
| initialize backbone from ALIGN instead of random (ablation) | OpenSeg (COCO+Loc. Narr.) · ADE20k-150 | mIoU: 18.6 -> 24.8 | Improved |
| initialize backbone from ALIGN instead of NoisyStudent (ablation) | OpenSeg (COCO+Loc. Narr.) · ADE20k-150 | mIoU: 24.4 -> 24.8 | No clear effect |
| filter captions to nouns and adjectives instead of all words (ablation) | OpenSeg (COCO) · ADE20k-150 / PASCAL-Context-59 | mIoU: 20.0 -> 21.1 (A-150); 41.3 -> 42.1 (PC-59) | Improved |
| randomly drop caption words with keep probability 0.75 instead of keeping all (ablation) | OpenSeg (COCO+Loc. Narr.) · ADE20k-150 | mIoU: 22.7 -> 24.8 | 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.
- GroupViT — GroupViT: Semantic Segmentation Emerges From Text Supervision (CVPR 2022) · better on 1 of 1 shared result
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)
- Exploring Simple Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.91)
- Training-Free Open-Vocabulary Segmentation with Offline Diffusion-Augmented Prototype Generation (CVPR 2024 · similarity 0.90)
- Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation (CVPR 2025 · similarity 0.90)
- Open-Vocabulary Segmentation with Semantic-Assisted Calibration (CVPR 2024 · similarity 0.89)
- Training-free Open-Vocabulary Semantic Segmentation via Diverse Prototype Construction and Sub-region Matching (AAAI 2025 · 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.
