List your company
ResearchFormerly WizWand

Open-Vocabulary Semantic Segmentation on Cityscapes (val)

75.1mIoUCONCEPTBANK
2.19621.12340.0558.977Oct 2, 2024Jan 17, 2025May 4, 2025Aug 19, 2025Dec 4, 2025Mar 21, 2026Jul 6, 2026
Updated 3mo ago

Evaluation results

MethodLinks
CONCEPTBANKBackbone (Size)=SAM3 (PE-L+/14), Optimization protocol=Parameter-free2026.0275.1
Oracleevaluation protocol=oracle2025.0367.47
SemLAfusion strategy=LoRA merging, tau=0.05, K=5, evaluation protocol=leave-one-out2025.0363.91
SemLAfusion strategy=Late Fusion2025.0362.86
SAM3Backbone (Size)=SAM3 (PE-L+/14), Optimization protocol=Parameter-free2026.0262.3
Uniform mergingfusion strategy=LoRA merging2025.0362.18
Uniformfusion strategy=Late Fusion2025.0361.82
ReMEBackbone (Size)=CLIP + DINOv2 (ViT-L/14) + SAM (ViT-L/16), Optimization protocol=Parameter-free2026.0259
T-REN+Approach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=576, Input Resolution=1344p2026.0458.7
T-RENApproach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=576, Input Resolution=384p2026.0452.7
CorrCLIPBackbone (Size)=CLIP (ViT-L/14) + DINO (ViT-B/8) + SAM2 (Hiera-L), Optimization protocol=Parameter-free2026.0251.1
TextRegionApproach Type=SAM-guided, Backbone=ViT-H, Input Resolution=1344p2026.0447.4
CAT-Segevaluation protocol=zero-shot2025.0347.11
TridentApproach Type=SAM-guided, Backbone=ViT-H, Input Resolution=1344p2026.0446.9
RADSeg+Approach Type=SAM-guided, Backbone=ViT-H, Input Resolution=1344p2026.0445.8
RF-CLIPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0243
TridentBackbone (Size)=CLIP + DINO (ViT-B/16) + SAM (ViT-B/16), Optimization protocol=Parameter-free2026.0242.9
SFPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0241.1
DIH-CLIPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0240.2
CASSBackbone (Size)=CLIP + DINO (ViT-B/8), Optimization protocol=Parameter-free2026.0239.4
Talk2DINOBackbone (Size)=CLIP + DINOv2 with registers (ViT-B/14), Optimization protocol=Training-based2026.0238.1
LPOSSBackbone (Size)=CLIP + DINO (ViT-B/16), Optimization protocol=Parameter-free2026.0237.3
CLIPixtraining=false2026.0737.1
DINOv3 dino.txtApproach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=5762026.0436.9
ResCLIPtraining=false2026.0735.9
NACLIPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0235.5
FreeDABackbone (Size)=CLIP + DINOv2 (ViT-B/14), Optimization protocol=Parameter-free2026.0234.1
ProxyCLIPBackbone (Size)=CLIP + DINOv2 with registers (ViT-B/14), Optimization protocol=Parameter-free2026.0233.9
SCLIPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0232.2
SCLIPtraining=false2026.0732.2
CLIP-DINOiserBackbone (Size)=CLIP + DINO (ViT-B/16), Optimization protocol=Training-based2026.0231.7
CLIP-DINOisertraining=true2026.0731.7
ClearCLIPUpsampling=SimFeatUp2024.1030.7
SCLIPUpsampling=SimFeatUp2024.1030.5
ClearCLIPUpsampling=Original2024.1030
ClearCLIPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0230
ClearCLIPtraining=false2026.0730
SCLIPUpsampling=Original2024.1029.1
CoDeBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Training-based2026.0228.9
DINOv2 dino.txtApproach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=5762026.0427.4
LaVGBackbone (Size)=CLIP + DINO (ViT-B/8), Optimization protocol=Parameter-free2026.0226.2
MaskCLIPUpsampling=SimFeatUp2024.1025.8
TCLUpsampling=Original2024.1023.1
TCLBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Training-based2026.0223.1
TCLtraining=true2026.0723.1
PEApproach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=5762026.0421.4
CLIPtraseBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.0221.3
RecoUpsampling=Original2024.1021.1
SigLIP-2Approach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=5762026.0416.3
EVA-02-CLIPApproach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=5762026.0414.1
MaskCLIPUpsampling=Original2024.1012.6
CLIPApproach Type=Patch-based vision-language encoder, Backbone=ViT-L, Tokens per image=5762026.0411.5
GroupViTBackbone (Size)=GroupViT (ViT-S/16), Optimization protocol=Training-based2026.0211.1
GroupViTtraining=true2026.0711.1
CLIPBackbone (Size)=CLIP (ViT-B/16), Optimization protocol=Parameter-free2026.026.7
CLIPtraining=false2026.075