Research map / Detection and segmentation / CNN features
OneFormer: One Transformer To Rule Universal Image 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 a task token to condition the model on the segmentation task (ablation) | OneFormer · Cityscapes val | AP: 43.3 (-2.3) -> 45.6 | Helped |
| Add learnable text context embeddings (ablation) | OneFormer · Cityscapes val | PQ: 62.7 (-4.5) -> 67.2 | Helped |
| Initialize queries by repeating the task token rather than using all zeros (ablation) | OneFormer · Cityscapes val | PQ: 65.8 (-1.4) -> 67.2 | Helped |
| Add query-text contrastive loss (ablation) | OneFormer · Cityscapes val | PQ: 58.8 (-8.4) -> 67.2 | Helped |
| Use query-text contrastive loss instead of query classification loss (ablation) | OneFormer · Cityscapes val | PQ: 66.4 (-0.8) -> 67.2 | Helped |
| Use the “a photo with a {CLS}” text template instead of “a photo with a {CLS} {TYPE}” (ablation) | OneFormer · Cityscapes val | PQ: 65.4 (-1.8) -> 67.2 | Helped |
| Use the “a photo with a {CLS}” text template instead of “{CLS}” (ablation) | OneFormer · Cityscapes val | PQ: 66.6 (-0.6) -> 67.2 | Helped |
| Use task-conditioned joint training rather than joint-training Mask2Former (ablation) | OneFormer · ADE20K val | AP: 33.7 (-2.2) -> 35.9 | Helped |
| Supply the panoptic task token rather than the instance task token (ablation) | OneFormer · ADE20K val | PQ: 33.1 -> 49.3 | Mixed |
| Supply the panoptic task token rather than the semantic task token (ablation) | OneFormer · ADE20K val | PQ: 40.4 -> 49.3 | Mixed |
| Train one jointly conditioned model rather than separately training task-specific models (ablation) | OneFormer · ADE20K val | PQ: 48.7 -> 49.8 | Helped |
Closest papers by meaning
- Unsupervised Universal Image Segmentation (CVPR 2024 · similarity 0.89)
- OneFormer3D: One Transformer for Unified Point Cloud Segmentation (CVPR 2024 · similarity 0.88)
- Segment Anything without Supervision (NeurIPS 2024 · similarity 0.88)
- OMG-Seg: Is One Model Good Enough For All Segmentation? (CVPR 2024 · similarity 0.87)
- Universal Segmentation at Arbitrary Granularity with Language Instruction (CVPR 2024 · similarity 0.87)
- Segment Anything, Even Occluded (CVPR 2025 · similarity 0.86)
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.