Masked-attention Mask Transformer for Universal Image Segmentation

Image segmentation is about grouping pixels with different semantics, e.g., category or instance membership, where each choice of semantics defines a task. While only the semantics of each task differ, current research focuses on designing specialized architectures for each task. We present Masked-attention Mask Transformer (Mask2Former), a new architecture capable of addressing any image segmentation task (panoptic, instance or semantic). Its key components include masked attention, which extracts localized features by constraining cross-attention within predicted mask regions. In addition to reducing the research effort by at least three times, it outperforms the best specialized architectures by a significant margin on four popular datasets. Most notably, Mask2Former sets a new state-of-the-art for panoptic segmentation (57.8 PQ on COCO), instance segmentation (50.1 AP on COCO) and semantic segmentation (57.7 mIoU on ADE20K).

PDF Abstract CVPR 2022 PDF CVPR 2022 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Semantic Segmentation ADE20K Mask2Former (SwinL-FaPN) Validation mIoU 57.7 # 18
Semantic Segmentation ADE20K Mask2Former (SwinL) Validation mIoU 57.3 # 23
Semantic Segmentation ADE20K Mask2Former (Swin-L-FaPN) Validation mIoU 56.4 # 28
Panoptic Segmentation ADE20K val Mask2Former (Swin-L + FAPN, 640x640) PQ 46.2 # 14
AP 33.2 # 12
mIoU 55.4 # 11
Panoptic Segmentation ADE20K val Mask2Former (Swin-L, 640x640) PQ 48.1 # 13
AP 34.2 # 11
mIoU 54.5 # 14
Semantic Segmentation ADE20K val Mask2Former (Swin-L-FaPN, multiscale) mIoU 57.7 # 14
Instance Segmentation ADE20K val Mask2Former (Swin-L + FAPN) AP 33.4 # 8
APS 14.6 # 5
APM 37.6 # 5
APL 54.6 # 5
Instance Segmentation ADE20K val Mask2Former (Swin-L, single-scale) AP 34.9 # 7
APS 16.3 # 3
APM 40 # 3
APL 54.7 # 4
Panoptic Segmentation ADE20K val Panoptic-DeepLab (SwideRNet) PQ 37.9 # 18
mIoU 50 # 15
Instance Segmentation ADE20K val Mask2Former (ResNet-50) APM 28.9 # 6
APL 43.1 # 6
Panoptic Segmentation ADE20K val Mask2Former (ResNet-50, 640x640) AP 26.5 # 13
mIoU 46.1 # 16
Semantic Segmentation ADE20K val Mask2Former (Swin-L-FaPN) mIoU 56.4 # 21
Instance Segmentation ADE20K val Mask2Former (ResNet50) AP 26.4 # 9
APS 10.4 # 6
Panoptic Segmentation ADE20K val Mask2Former (ResNet50, 640x640) PQ 39.7 # 17
Instance Segmentation Cityscapes val Mask2Former (Swin-B) mask AP 42 # 5
Semantic Segmentation Cityscapes val Mask2Former (Swin-L) mIoU 84.3 # 10
Panoptic Segmentation Cityscapes val Mask2Former (Swin-L) PQ 66.6 # 12
mIoU 82.9 # 12
AP 43.6 # 10
Instance Segmentation Cityscapes val Mask2Former (Swin-S) mask AP 41.8 # 6
Instance Segmentation Cityscapes val Mask2Former (Swin-T) mask AP 39.7 # 8
Instance Segmentation Cityscapes val Mask2Former (ResNet-101) mask AP 38.5 # 9
Instance Segmentation Cityscapes val Mask2Former (ResNet-50) mask AP 37.4 # 10
Instance Segmentation Cityscapes val Mask2Former (Swin-L, single-scale) mask AP 43.7 # 4
Semantic Segmentation COCO MaskFormer (Swin-L, single-scale) mIoU 64.8 # 4
Semantic Segmentation COCO Mask2Former (Swin-L, single-scale) mIoU 67.4 # 2
Instance Segmentation COCO minival Mask2Former (Swin-L) mask AP 50.1 # 17
Panoptic Segmentation COCO minival Mask2Former (single-scale) PQ 57.8 # 11
PQth 64.2 # 6
PQst 48.1 # 7
AP 48.6 # 6
Panoptic Segmentation COCO test-dev Mask2Former (Swin-L) PQ 58.3 # 3
PQst 48.1 # 2
PQth 65.1 # 1
Instance Segmentation COCO test-dev Mask2Former (Swin-L, single scale) mask AP 50.5 # 14
AP50 74.9 # 3
AP75 54.9 # 3
APS 29.1 # 7
APM 53.8 # 2
APL 71.2 # 2
Instance Segmentation COCO val (panoptic labels) Mask2Former (Swin-L, single-scale) AP 49.1 # 2
Semantic Segmentation Mapillary val Mask2Former (Swin-L, multiscale) mIoU 64.7 # 2

Methods