mask2former

Introduction

Official Repo

Code Snippet

Mask2Former (CVPR'2022)

@inproceedings{cheng2021mask2former,
    title={Masked-attention Mask Transformer for Universal Image Segmentation},
    author={Bowen Cheng and Ishan Misra and Alexander G. Schwing and Alexander Kirillov and Rohit Girdhar},
    journal={CVPR},
    year={2022}
}

Results

ADE20k

Segmentor	Pretrain	Backbone	Crop Size	Schedule	Train/Eval Set	mIoU	Download
Mask2Former	ImageNet-1k-224x224	Swin-T	512x512	LR/POLICY/BS/EPOCH: 0.0001/poly/16/130	train/val	49.20%	cfg | model | log
Mask2Former	ImageNet-1k-224x224	Swin-S	512x512	LR/POLICY/BS/EPOCH: 0.0001/poly/16/130	train/val	51.37%	cfg | model | log
Mask2Former	ImageNet-22k-384x384	Swin-B	640x640	LR/POLICY/BS/EPOCH: 0.0001/poly/16/130	train/val	54.13%	cfg | model | log
Mask2Former	ImageNet-22k-384x384	Swin-L	640x640	LR/POLICY/BS/EPOCH: 0.0001/poly/16/130	train/val	56.30%	cfg | model | log

Cityscapes

Segmentor	Pretrain	Backbone	Crop Size	Schedule	Train/Eval Set	mIoU	Download
Mask2Former	ImageNet-1k-224x224	Swin-T	512x1024	LR/POLICY/BS/EPOCH: 0.0001/poly/16/500	train/val	82.10%	cfg | model | log
Mask2Former	ImageNet-1k-224x224	Swin-S	512x1024	LR/POLICY/BS/EPOCH: 0.0001/poly/16/500	train/val	82.65%	cfg | model | log
Mask2Former	ImageNet-22k-384x384	Swin-B	512x1024	LR/POLICY/BS/EPOCH: 0.0001/poly/16/500	train/val	83.62%	cfg | model | log
Mask2Former	ImageNet-22k-384x384	Swin-L	512x1024	LR/POLICY/BS/EPOCH: 0.0001/poly/16/500	train/val	83.79%	cfg | model | log

More

You can also download the model weights from following sources:

• BaiduNetdisk: https://pan.baidu.com/s/1gD-NJJWOtaHCtB0qHE79rA with access code s757