Instance Segmentation on MS COCO (mAP Mask)
47.9mAP MaskInfoMamba-B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| InfoMamba-BParams (M)=145, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 47.9 | 70.7 | 52.2 | |
| InfoMamba-SParams (M)=108, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 47.4 | 70.5 | 51.7 | |
| InfoMamba-TParams (M)=86, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 46.5 | 69.3 | 50.7 | |
| MAELoss Function=dVAE-P, Pre-training Data=IN1K + DALL-E2022.12 | 46.4 | — | — | |
| MSG-MAELoss Function=StyleGANv2-ADA-P, Pre-training Data=IN1K2022.12 | 46.1 | — | — | |
| MAELoss Function=MSG-GAN-P, Pre-training Data=IN1K2022.12 | 45.9 | — | — | |
| MSG-MAELoss Function=MSG-GAN-P, Pre-training Data=IN1K2022.12 | 45.8 | — | — | |
| DSeq-C-JEPAArch.=ViT-B/16, Epochs=600, View data augmentations=without2025.11 | 45.7 | — | — | |
| MambaVision-BParams (M)=145, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 45.7 | 68.7 | 49.4 | |
| ConvNeXt-BParams (M)=146, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 45.6 | 68.9 | 49.5 | |
| Deformba-BSchedule=3x MS, Resolution=1280 x 800, #Param.=106M, FLOPs=496G2026.05 | 45.6 | 69.8 | 49.3 | |
| MAELoss Function=StyleGANv2-ADA-P, Pre-training Data=IN1K2022.12 | 45.5 | — | — | |
| MAELoss Function=LS-GAN-P, Pre-training Data=IN1K2022.12 | 45.4 | — | — | |
| C-JEPAArch.=ViT-B/16, Epochs=600, View data augmentations=without2025.11 | 45.3 | — | — | |
| MambaVision-SParams (M)=108, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 45.2 | 68.5 | 48.9 | |
| Deformba-BSchedule=1x, Resolution=1280 x 800, #Param.=106M, FLOPs=496G2026.05 | 45.2 | 69.3 | 48.9 | |
| MAELoss Function=MS-SSIM + L1, Pre-training Data=IN1K2022.12 | 45.1 | — | — | |
| Spatial-Mamba-BSchedule=1x, Resolution=1280 x 800, #Param.=115M, FLOPs=494G2026.05 | 45.1 | 69.1 | 49.1 | |
| VSSD-SSchedule=3x MS, Resolution=1280 x 800, #Param.=77M, FLOPs=381G2026.05 | 45.1 | 69.4 | 48.8 | |
| Deformba-SSchedule=3x MS, Resolution=1280 x 800, #Param.=64M, FLOPs=374G2026.05 | 45.1 | 69.2 | 48.9 | |
| DSeq-JEPAArch.=ViT-B/16, Epochs=600, View data augmentations=without2025.11 | 45 | — | — | |
| Swin-SParams (M)=107, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 45 | 68.2 | 48.8 | |
| ConvNeXt-SParams (M)=108, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 45 | 68.4 | 49.1 | |
| Swin-BParams (M)=145, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 45 | 68.1 | 48.9 | |
| MAELoss Function=MSE, Pre-training Data=IN1K2022.12 | 44.9 | — | — | |
| MAEArch.=ViT-B/16, Epochs=1600, View data augmentations=without2025.11 | 44.9 | — | — | |
| VSSD-BSchedule=1x, Resolution=1280 x 800, #Param.=108M, FLOPs=506G2026.05 | 44.8 | 69.1 | 48.3 | |
| Deformba-SSchedule=1x, Resolution=1280 x 800, #Param.=64M, FLOPs=374G2026.05 | 44.7 | 68.7 | 48.3 | |
| Spatial-Mamba-SSchedule=3x MS, Resolution=1280 x 800, #Param.=63M, FLOPs=315G2026.05 | 44.6 | 68.7 | 68.7 | |
| I-JEPAArch.=ViT-B/16, Epochs=600, View data augmentations=without2025.11 | 44.5 | — | — | |
| Deformba-TSchedule=3x MS, Resolution=1280 x 800, #Param.=45M, FLOPs=266G2026.05 | 44.5 | 68.5 | 48.2 | |
| VNCT-TTraining Schedule=3x + MS2026.07 | 44.5 | 68.6 | 47.9 | |
| BEITLoss Function=Negative Log Likelihood, Pre-training Data=IN1K + DALL-E2022.12 | 44.4 | — | — | |
| MambaVision-TParams (M)=86, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 44.3 | 67.3 | 47.9 | |
| iBOTArch.=ViT-B/16, Epochs=1600, View data augmentations=with2025.11 | 44.2 | — | — | |
| MSVMamba-SSchedule=3x MS, Resolution=1280 x 800, #Param.=70M, FLOPs=349G2026.05 | 44.2 | 68 | 47.9 | |
| VMamba-SSchedule=3x MS, Resolution=1280 x 800, #Param.=70M, FLOPs=349G2026.05 | 44.2 | 68.2 | 47.7 | |
| VMamba-BBackbone=VMamba-B, #Param=108M, FLOPs=485G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 44.1 | 68.3 | 47.7 | |
| LocalVMamba-SSchedule=3x MS, Resolution=1280 x 800, #Param.=69M, FLOPs=414G2026.05 | 44.1 | 67.8 | 47.4 | |
| Spatial-Mamba-SSchedule=1x, Resolution=1280 x 800, #Param.=63M, FLOPs=315G2026.05 | 44 | 67.9 | 47.5 | |
| VSSD-TSchedule=3x MS, Resolution=1280 x 800, #Param.=52M, FLOPs=298G2026.05 | 44 | 67.6 | 47.4 | |
| VSSD-SSchedule=1x, Resolution=1280 x 800, #Param.=77M, FLOPs=381G2026.05 | 43.9 | 68.1 | 47.4 | |
| CSWin-BSchedule=1x, Resolution=1280 x 800, #Param.=88M, FLOPs=496G2026.05 | 43.9 | 67.8 | 47.3 | |
| VMamba-BSchedule=1x, Resolution=1280 x 800, #Param.=108M, FLOPs=485G2026.05 | 43.9 | 67.7 | 47.6 | |
| MFilMamba-BBackbone=MFilMamba-B, #Param=112M, FLOPs=467G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 43.8 | 67.7 | 47.6 | |
| Spatial-Mamba-TSchedule=3x MS, Resolution=1280 x 800, #Param.=46M, FLOPs=261G2026.05 | 43.8 | 67.8 | 47.2 | |
| Swin-TParams (M)=86, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 43.7 | 66.6 | 47.3 | |
| ConvNeXt-TParams (M)=86, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 43.7 | 66.5 | 47.3 | |
| VMamba-SBackbone=VMamba-S, #Param=70M, FLOPs=349G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 43.7 | 67.3 | 47 | |
| VMamba-SSchedule=1x, Resolution=1280 x 800, #Param.=70M, FLOPs=349G2026.05 | 43.7 | 67.3 | 47 | |
| VMamba-TSchedule=3x MS, Resolution=1280 x 800, #Param.=50M, FLOPs=271G2026.05 | 43.7 | 67.4 | 47 | |
| VSSD-TTraining Schedule=3x + MS2026.07 | 43.6 | 67.6 | 46.9 | |
| ConvNeXt-BSchedule=3x MS, Resolution=1280 x 800, #Param.=108M, FLOPs=486G2026.05 | 43.5 | 67.1 | 46.7 | |
| VNCT-TTraining Schedule=1x2026.07 | 43.5 | 67.3 | 46.8 | |
| DINOArch.=ViT-B/16, Epochs=1600, View data augmentations=with2025.11 | 43.4 | — | — | |
| MSVMamba-TSchedule=3x MS, Resolution=1280 x 800, #Param.=52M, FLOPs=275G2026.05 | 43.4 | 67.2 | 46.8 | |
| LocalVMamba-TSchedule=3x MS, Resolution=1280 x 800, #Param.=45M, FLOPs=291G2026.05 | 43.4 | 67 | 46.4 | |
| LocalVMamba-TTraining Schedule=3x + MS2026.07 | 43.4 | 67 | 46.4 | |
| Deformba-TSchedule=1x, Resolution=1280 x 800, #Param.=45M, FLOPs=266G2026.05 | 43.3 | 66.8 | 46.5 | |
| Swin-BSchedule=3x MS, Resolution=1280 x 800, #Param.=107M, FLOPs=496G2026.05 | 43.3 | 67.1 | 46.7 | |
| CSWin-SSchedule=1x, Resolution=1280 x 800, #Param.=54M, FLOPs=342G2026.05 | 43.2 | 67.1 | 46.2 | |
| MSVMamba-SSchedule=1x, Resolution=1280 x 800, #Param.=70M, FLOPs=349G2026.05 | 43.2 | 67.3 | 46.5 | |
| Swin-SSchedule=3x MS, Resolution=1280 x 800, #Param.=69M, FLOPs=354G2026.05 | 43.2 | 67 | 46.1 | |
| PVTv2-B3Schedule=3x MS, Resolution=1280 x 800, #Param.=65M, FLOPs=397G2026.05 | 43.2 | 66.9 | 46.7 | |
| VMamba-TTraining Schedule=3x + MS2026.07 | 43.2 | 66.8 | 46.3 | |
| MFilMamba-SBackbone=MFilMamba-S, #Param=70M, FLOPs=320G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 43.1 | 66.8 | 46.4 | |
| PVTv2-B2Schedule=3x MS, Resolution=1280 x 800, #Param.=45M, FLOPs=309G2026.05 | 43.1 | 66.8 | 46.7 | |
| MambaOut-BaseBackbone=MambaOut-Base, #Param=100M, FLOPs=495G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 43 | 66.4 | 46.3 | |
| Spatial-Mamba-TSchedule=1x, Resolution=1280 x 800, #Param.=46M, FLOPs=261G2026.05 | 42.9 | 66.5 | 46.2 | |
| ConvNeXt-SSchedule=3x MS, Resolution=1280 x 800, #Param.=70M, FLOPs=348G2026.05 | 42.9 | 66.9 | 46.2 | |
| PVTv2-B5Schedule=3x MS, Resolution=1280 x 800, #Param.=102M, FLOPs=557G2026.05 | 42.9 | 66.6 | 46.2 | |
| VSSD-TSchedule=1x, Resolution=1280 x 800, #Param.=52M, FLOPs=298G2026.05 | 42.8 | 66.5 | 46.1 | |
| DefMamba-SSchedule=1x, Resolution=1280 x 800, FLOPs=268G2026.05 | 42.8 | 66.3 | 46.2 | |
| MoCo v3Loss Function=InfoNCE [42], Pre-training Data=IN1K2022.12 | 42.7 | — | — | |
| VMamba-TBackbone=VMamba-T, #Param=50M, FLOPs=271G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.7 | 66.4 | 45.9 | |
| MFilMamba-TBackbone=MFilMamba-T, #Param=53M, FLOPs=261G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.7 | 66.2 | 46 | |
| MambaOut-SmallBackbone=MambaOut-Small, #Param=65M, FLOPs=354G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.7 | 66.1 | 46.2 | |
| ConvNeXt-BBackbone=ConvNeXt-B, #Param=108M, FLOPs=486G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.7 | 66.3 | 46 | |
| VMamba-TSchedule=1x, Resolution=1280 x 800, #Param.=50M, FLOPs=271G2026.05 | 42.7 | 66.4 | 45.9 | |
| MambaOut-SSchedule=1x, Resolution=1280 x 800, #Param.=65M, FLOPs=354G2026.05 | 42.7 | 66.1 | 46.2 | |
| ConvNeXt-BSchedule=1x, Resolution=1280 x 800, #Param.=111M, FLOPs=507G2026.05 | 42.7 | 66.3 | 46 | |
| VSSD-TTraining Schedule=1x2026.07 | 42.6 | 66.4 | 45.9 | |
| DAT-SBackbone=DAT-S, #Param=69M, FLOPs=378G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.5 | 66.7 | 45.4 | |
| MSVMamba-TSchedule=1x, Resolution=1280 x 800, #Param.=52M, FLOPs=275G2026.05 | 42.5 | 66.2 | 45.8 | |
| DAT-SSchedule=1x, Resolution=1280 x 800, #Param.=69M, FLOPs=378G2026.05 | 42.5 | 66.7 | 45.4 | |
| FractalMamba-TBackbone=FractalMamba-T, #Param=41M, FLOPs=266G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.4 | 65.9 | 45.8 | |
| DAT-TSchedule=1x, Resolution=1280 x 800, #Param.=48M, FLOPs=272G2026.05 | 42.4 | 66.1 | 45.5 | |
| Swin-BBackbone=Swin-B, #Param=107M, FLOPs=496G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 42.3 | — | — | |
| Swin-BSchedule=1x, Resolution=1280 x 800, #Param.=107M, FLOPs=496G2026.05 | 42.3 | 66 | 45.5 | |
| CSWin-TSchedule=1x, Resolution=1280 x 800, #Param.=42M, FLOPs=279G2026.05 | 42.2 | 65.6 | 45.4 | |
| LocalVMamba-TTraining Schedule=1x2026.07 | 42.2 | 65.7 | 45.5 | |
| VMamba-TTraining Schedule=1x2026.07 | 42.1 | 65.5 | 45.3 | |
| ConvNeXt-SBackbone=ConvNeXt-S, #Param=70M, FLOPs=348G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 41.8 | 65.2 | 45.1 | |
| ViT-Adapter-BBackbone=ViT-Adapter-B, #Param=120M, FLOPs=557G, Framework=Mask R-CNN, Training schedule=1×, Input resolution=1280 × 800, Pre-training=ImageNet-1K2026.03 | 41.8 | 65.1 | 44.9 | |
| ConvNeXt-SSchedule=1x, Resolution=1280 x 800, #Param.=70M, FLOPs=348G2026.05 | 41.8 | 65.2 | 45.1 | |
| ViT-Adapter-BSchedule=1x, Resolution=1280 x 800, #Param.=102M, FLOPs=557G2026.05 | 41.8 | 65.1 | 44.9 | |
| X101-64Params (M)=140, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 41.7 | 64 | 45.1 | |
| ConvNeXt-TSchedule=3x MS, Resolution=1280 x 800, #Param.=48M, FLOPs=262G2026.05 | 41.7 | 65 | 44.9 | |
| ConvNeXt-TTraining Schedule=3x + MS2026.07 | 41.7 | 65 | 44.9 | |
| X101-32Params (M)=101, Detector=Cascade Mask R-CNN, Schedule=3×, Crop Size=1280 × 8002026.03 | 41.6 | 63.9 | 45.2 |