Masked-attention Mask Transformer for Universal Image Segmentation

2 Dec 2021  ·  Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, Rohit Girdhar ·

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). read more

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Semantic Segmentation ADE20K val Mask2Former (Swin-L-FaPN) mIoU 56.4 # 9
Semantic Segmentation ADE20K val Mask2Former (Swin-L-FaPN, multiscale) mIoU 57.7 # 3
Panoptic Segmentation COCO minival Mask2Former (single-scale) PQ 57.8 # 1
PQth 64.2 # 1
PQst 48.1 # 1
Instance Segmentation COCO minival Mask2Former (Swin-L) mask AP 50.1 # 7
APL 72.1 # 1
APM 53.9 # 3
APS 29.9 # 4
Panoptic Segmentation COCO test-dev Mask2Former (Swin-L) PQ 58.3 # 1
PQst 48.1 # 1
PQth 65.1 # 1
Instance Segmentation COCO test-dev Mask2Former (Swin-L) mask AP 50.5 # 7
AP50 74.9 # 2
AP75 54.9 # 2
APS 29.1 # 6
APM 53.8 # 1
APL 71.2 # 1

Methods