Your ViT is Secretly an Image Segmentation Model

Benchmark Model Rank Results
panoptic-segmentation-on-ade20k-valEoMT (DINOv2-g, single-scale, 1280x1280, COCO pre-trained)#5PQ: 52.8
panoptic-segmentation-on-coco-minivalEoMT (DINOv2-g, single-scale, 1280x1280)#7PQ: 59.2
semantic-segmentation-on-ade20k-valEoMT (DINOv2-L, single-scale, 512x512)#12mIoU: 58.4
semantic-segmentation-on-cityscapes-valEoMT (DINOv2-L, single-scale, 1024x1024)#18mIoU: 84.2FPS: 25Validation mIoU: 84.2
semantic-segmentation-on-ade20kEoMT (DINOv2-L, single-scale, 512x512)#23Validation mIoU: 58.4Params (M): 316GFLOPs (512 x 512): 721…
instance-segmentation-on-coco-minivalEoMT (ViT-L)#37mask AP: 48.8