Dilated Neighborhood Attention Transformer

Open paper
Benchmark Model Rank Results
instance-segmentation-on-cityscapes-valDiNAT-L (single-scale, Mask2Former)#7mask AP: 45.1AP50: 72.6
panoptic-segmentation-on-coco-minivalDiNAT-L (single-scale, Mask2Former)#9PQ: 58.5PQst: 48.8PQth: 64.9AP: 49.2mIoU: 68.3
instance-segmentation-on-ade20k-valDiNAT-L (Mask2Former, single-scale)#10AP: 35.4APL: 55.5APM: 39.0APS: 16.3
panoptic-segmentation-on-cityscapes-valDiNAT-L (Mask2Former)#10PQ: 67.2mIoU: 83.4AP: 44.5
semantic-segmentation-on-cityscapes-valDiNAT-L (Mask2Former)#14mIoU: 84.5
panoptic-segmentation-on-ade20k-valDiNAT-L (Mask2Former, 640x640)#16PQ: 49.4mIoU: 56.3AP: 35.0
semantic-segmentation-on-ade20k-valDiNAT-L (Mask2Former)#16mIoU: 58.1
instance-segmentation-on-coco-minivalDiNAT-L (single-scale, Mask2Former)#21mask AP: 50.8AP50: 75.0
semantic-segmentation-on-ade20kDiNAT-L (Mask2Former)#27Validation mIoU: 58.1
semantic-segmentation-on-ade20kDiNAT-Large (UperNet)#54Validation mIoU: 54.9
semantic-segmentation-on-ade20kDiNAT_s-Large (UperNet)#61Validation mIoU: 54.6
semantic-segmentation-on-ade20kDiNAT-Base (UperNet)#123Validation mIoU: 50.4
semantic-segmentation-on-ade20kDiNAT-Small (UperNet)#134Validation mIoU: 49.9
semantic-segmentation-on-ade20kDiNAT-Tiny (UperNet)#157Validation mIoU: 48.8
semantic-segmentation-on-ade20kDiNAT-Mini (UperNet)#182Validation mIoU: 47.2
image-classification-on-imagenetDiNAT-Large (11x11ks; 384res; Pretrained on IN22K@224)#90Top 1 Accuracy: 87.5%Number of params: 200MGFLOPs: 92.4
image-classification-on-imagenetDiNAT-Large (384x384; Pretrained on ImageNet-22K @ 224x224)#97Top 1 Accuracy: 87.4%GFLOPs: 89.7
image-classification-on-imagenetDiNAT_s-Large (384res; Pretrained on IN22K@224)#98Top 1 Accuracy: 87.4%Number of params: 197MGFLOPs: 101.5
image-classification-on-imagenetDiNAT_s-Large (224x224; Pretrained on ImageNet-22K @ 224x224)#149Top 1 Accuracy: 86.5%GFLOPs: 34.5
image-classification-on-imagenetDiNAT-Base#345Top 1 Accuracy: 84.4%Number of params: 90MGFLOPs: 13.7
image-classification-on-imagenetDiNAT-Small#419Top 1 Accuracy: 83.8%Number of params: 51MGFLOPs: 7.8
image-classification-on-imagenetDiNAT-Tiny#545Top 1 Accuracy: 82.7%Number of params: 28MGFLOPs: 4.3
image-classification-on-imagenetDiNAT-Mini#645Top 1 Accuracy: 81.8%Number of params: 20MGFLOPs: 2.7
object-detection-on-coco-val2017DiNAT-L + Cascade Mask R-CNN#102box AP: 55.3

Rank counts only results with a code link.