kMaX-DeepLab: k-means Mask Transformer

Benchmark Model Rank Results
panoptic-segmentation-on-ade20k-valkMaX-DeepLab (ConvNeXt-L, single-scale, 1281x1281)#10PQ: 50.9mIoU: 55.2
panoptic-segmentation-on-ade20k-valkMaX-DeepLab (ConvNeXt-L, single-scale, 641x641)#17PQ: 48.7mIoU: 54.8
panoptic-segmentation-on-ade20k-valkMaX-DeepLab (ResNet50, single-scale, 1281x1281)#20PQ: 42.3mIoU: 45.3
panoptic-segmentation-on-ade20k-valkMaX-DeepLab (ResNet50, single-scale, 641x641)#21PQ: 41.5mIoU: 45.0
panoptic-segmentation-on-cityscapes-testkMaX-DeepLab (single-scale)#4PQ: 66.2
panoptic-segmentation-on-cityscapes-valkMaX-DeepLab (single-scale)#5PQ: 68.4mIoU: 83.5AP: 44.0
panoptic-segmentation-on-coco-minivalkMaX-DeepLab (single-scale, pseudo-labels)#11PQ: 58.1PQst: 48.8PQth: 64.3
panoptic-segmentation-on-coco-minivalkMaX-DeepLab (single-scale, drop query with 256 queries)#13PQ: 58.0PQst: 48.6PQth: 64.2
panoptic-segmentation-on-coco-minivalkMaX-DeepLab (single-scale)#16PQ: 57.9PQst: 48.6PQth: 64.0
panoptic-segmentation-on-coco-test-devkMaX-DeepLab (single-scale)#2PQ: 58.5PQst: 49.0PQth: 64.8
semantic-segmentation-on-cityscapeskMaX-DeepLab (ConvNeXt-L, fine only)#16Mean IoU (class): 83.2%