Home

Awesome

SeMask: Semantically Masked Transformers

Framework: PyTorch License

Jitesh Jain, Anukriti Singh, Nikita Orlov, Zilong Huang, Jiachen Li, Steven Walton, Humphrey Shi

[arXiv] [pdf] [BibTeX]

This repo contains the code for our paper SeMask: Semantically Masked Transformers for Semantic Segmentation.

<img src="images/semask.svg" alt='semask' height='600px'>

Contents

  1. Results
  2. Setup Instructions
  3. Citing SeMask

1. Results

Note: † denotes the backbones were pretrained on ImageNet-22k and 384x384 resolution images.

ADE20K

MethodBackboneCrop SizemIoUmIoU (ms+flip)#paramsconfigCheckpoint
SeMask-T FPNSeMask Swin-T512x51242.0643.3635Mconfigcheckpoint
SeMask-S FPNSeMask Swin-S512x51245.9247.6356Mconfigcheckpoint
SeMask-B FPNSeMask Swin-B<sup></sup>512x51249.3550.9896Mconfigcheckpoint
SeMask-L FPNSeMask Swin-L<sup></sup>640x64051.8953.52211Mconfigcheckpoint
SeMask-L MaskFormerSeMask Swin-L<sup></sup>640x64054.7556.15219Mconfigcheckpoint
SeMask-L Mask2FormerSeMask Swin-L<sup></sup>640x64056.4157.52222Mconfigcheckpoint
SeMask-L Mask2Former FaPNSeMask Swin-L<sup></sup>640x64056.8858.25227Mconfigcheckpoint
SeMask-L Mask2Former MSFaPNSeMask Swin-L<sup></sup>640x64057.0058.25224Mconfigcheckpoint

Cityscapes

MethodBackboneCrop SizemIoUmIoU (ms+flip)#paramsconfigCheckpoint
SeMask-T FPNSeMask Swin-T768x76874.9276.5634Mconfigcheckpoint
SeMask-S FPNSeMask Swin-S768x76877.1379.1456Mconfigcheckpoint
SeMask-B FPNSeMask Swin-B<sup></sup>768x76877.7079.7396Mconfigcheckpoint
SeMask-L FPNSeMask Swin-L<sup></sup>768x76878.5380.39211Mconfigcheckpoint
SeMask-L Mask2FormerSeMask Swin-L<sup></sup>512x102483.9784.98222Mconfigcheckpoint

COCO-Stuff 10k

MethodBackboneCrop SizemIoUmIoU (ms+flip)#paramsconfigCheckpoint
SeMask-T FPNSeMask Swin-T512x51237.5338.8835Mconfigcheckpoint
SeMask-S FPNSeMask Swin-S512x51240.7242.2756Mconfigcheckpoint
SeMask-B FPNSeMask Swin-B<sup></sup>512x51244.6346.3096Mconfigcheckpoint
SeMask-L FPNSeMask Swin-L<sup></sup>640x64047.4748.54211Mconfigcheckpoint
<img src="SeMask-FPN/docs/demo.svg" alt='demo' height='600px'>

2. Setup Instructions

We provide the codebase with SeMask incorporated into various models. Please check the setup instructions inside the corresponding folders:

3. Citing SeMask

@inproceedings{jain2023semask,
title={SeMask: Semantically Masked Transformers for Semantic Segmentation}, 
author={Jitesh Jain and Anukriti Singh and Nikita Orlov and Zilong Huang and Jiachen Li and Steven Walton and Humphrey Shi},
year={2023},
booktitle={ICCV Workshops 2023},
}

Acknowledgements

Code is based heavily on the following repositories: Swin-Transformer-Semantic-Segmentation, Mask2Former, MaskFormer and FaPN-full.