Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective

Benchmark Model Rank Results
semantic-segmentation-on-pascal-contextDEPICT-SA (ViT-L multi-scale)#14mIoU: 58.6
semantic-segmentation-on-pascal-contextDEPICT-SA (ViT-L single-scale)#16mIoU: 57.9
semantic-segmentation-on-ade20k-valDEPICT-SA (ViT-L 640x640 multi-scale)#33mIoU: 54.3
semantic-segmentation-on-ade20k-valDEPICT-SA (ViT-L 640x640 single-scale)#41mIoU: 52.9
semantic-segmentation-on-cityscapes-valDEPICT-SA (ViT-L multi-scale)#43mIoU: 81.0
semantic-segmentation-on-cityscapes-valDEPICT-SA (ViT-L single-scale)#55mIoU: 78.8