Semantic Segmentation on COCO (Resolution-Scaled mIoU)
60.5mIoU (1/512 ratio)CLIP-Guided SAM (labelled-only)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CLIP-Guided SAM (labelled-only)Unlabelled data=No, Image-level prompts=Yes, Encoder=CLIP-B+SAM-B, Number of Parameters=197M2026.05 | 60.5 | 63.2 | 65.8 | 67.7 | 69.2 | — | |
| HVLFormerBackbone=SigLIP22025.06 | 59.4 | 61.3 | 63.7 | 65.1 | 67.4 | — | |
| HVLFormerBackbone=ViT-B2025.06 | 54.9 | 56.7 | 61.2 | 62.9 | 65.1 | — | |
| HVLFormerBackbone=E.02-S2025.06 | 52.7 | 54.8 | 56.2 | 58.5 | 61.1 | — | |
| SemiVLBackbone=ViT-B2025.06 | 50.1 | 52.8 | 53.6 | 55.4 | 56.5 | — | |
| SemiVLUnlabelled data=Yes, Image-level prompts=No, Encoder=CLIP-Base, Number of Parameters=88M2026.05 | 50.1 | 52.8 | 53.6 | 55.4 | 56.5 | — | |
| UniM V2Backbone=D.v2-B2025.06 | 47.9 | 55.8 | 58.7 | 60.4 | 63.3 | — | |
| UniMatchV2Unlabelled data=Yes, Image-level prompts=No, Encoder=DINOv2-B, Number of Parameters=97.5M2026.05 | 47.9 | 55.8 | 58.7 | 60.4 | 63.3 | — | |
| POSBackbone=D.v2-S2025.06 | 40.9 | 46.8 | 54.3 | 56.2 | 57.7 | — | |
| TQDMBackbone=SigLIP22025.06 | 40.2 | 48.7 | 54.3 | 58.9 | 62.7 | — | |
| Pseudo-SDBackbone=ViT-B2025.06 | 39.9 | 44.8 | 48.2 | 50.2 | 51.9 | — | |
| UniMatchV2 (labelled-only)Unlabelled data=No, Image-level prompts=No, Encoder=DINOv2-B, Number of Parameters=97.5M2026.05 | 36.8 | 45.8 | 52.1 | 56.2 | 59.5 | — | |
| UniM V1†Backbone=ViT-B2025.06 | 36.6 | 44.1 | 49.1 | 53.5 | 55 | — | |
| LogicD.Backbone=XC652025.06 | 33.1 | 40.3 | 45.4 | 48.8 | 50.5 | — | |
| UniM V1Backbone=XC652025.06 | 31.9 | 38.9 | 44.4 | 48.2 | 49.8 | — | |
| CLIP-Guided SAMFine-tuning SAM?=Yes, VLM Size=87M (CS-CLIP-Base)2026.05 | — | — | — | — | — | 60.5 | |
| CLIPSurgery+SAMFine-tuning SAM?=Yes (decoder), VLM Size=87M (CS-CLIP-Base)2026.05 | — | — | — | — | — | 34.6 | |
| CLIPSurgery+SAMFine-tuning SAM?=Yes (adapters + decoder), VLM Size=87M (CS-CLIP-Base)2026.05 | — | — | — | — | — | 36.2 | |
| CLIPSurgery+SAM (modified)Fine-tuning SAM?=No, VLM Size=87M (CS-CLIP-Base)2026.05 | — | — | — | — | — | 28 | |
| CLIPSurgery+SAM (vanilla)Fine-tuning SAM?=No, VLM Size=87M (CS-CLIP-Base)2026.05 | — | — | — | — | — | 23.9 | |
| GroundedSAMFine-tuning SAM?=No, VLM Size=172M (GroundingDINO)2026.05 | — | — | — | — | — | 61.3 |