Skip to content

Latest commit

 

History

History

setr

SETR

Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers

Introduction

Official Repo

Code Snippet

Abstract

Most recent semantic segmentation methods adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract/semantic visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated/atrous convolutions or inserting attention modules. However, the encoder-decoder based FCN architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating semantic segmentation as a sequence-to-sequence prediction task. Specifically, we deploy a pure transformer (ie, without convolution and resolution reduction) to encode an image as a sequence of patches. With the global context modeled in every layer of the transformer, this encoder can be combined with a simple decoder to provide a powerful segmentation model, termed SEgmentation TRansformer (SETR). Extensive experiments show that SETR achieves new state of the art on ADE20K (50.28% mIoU), Pascal Context (55.83% mIoU) and competitive results on Cityscapes. Particularly, we achieve the first position in the highly competitive ADE20K test server leaderboard on the day of submission.

This head has two version head.

Usage

You can download the pretrain from here. Then you can convert its keys with the script vit2mmseg.py in the tools directory.

python tools/model_converters/vit2mmseg.py ${PRETRAIN_PATH} ${STORE_PATH}

E.g.

python tools/model_converters/vit2mmseg.py \
jx_vit_large_p16_384-b3be5167.pth pretrain/vit_large_p16.pth

This script convert the model from PRETRAIN_PATH and store the converted model in STORE_PATH.

Results and models

ADE20K

Method Backbone Crop Size Batch Size Lr schd Mem (GB) Inf time (fps) Device mIoU mIoU(ms+flip) config download
SETR Naive ViT-L 512x512 16 160000 18.40 4.72 V100 48.28 49.56 config model | log
SETR PUP ViT-L 512x512 16 160000 19.54 4.50 V100 48.24 49.99 config model | log
SETR MLA ViT-L 512x512 8 160000 10.96 - V100 47.34 49.05 config model | log
SETR MLA ViT-L 512x512 16 160000 17.30 5.25 V100 47.39 49.37 config model | log

Cityscapes

Method Backbone Crop Size Batch Size Lr schd Mem (GB) Inf time (fps) Device mIoU mIoU(ms+flip) config download
SETR Naive ViT-L 768x768 8 80000 24.06 0.39 V100 78.10 80.22 config model | log
SETR PUP ViT-L 768x768 8 80000 27.96 0.37 V100 79.21 81.02 config model | log
SETR MLA ViT-L 768x768 8 80000 24.10 0.41 V100 77.00 79.59 config model | log

Citation

@article{zheng2020rethinking,
  title={Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers},
  author={Zheng, Sixiao and Lu, Jiachen and Zhao, Hengshuang and Zhu, Xiatian and Luo, Zekun and Wang, Yabiao and Fu, Yanwei and Feng, Jianfeng and Xiang, Tao and Torr, Philip HS and others},
  journal={arXiv preprint arXiv:2012.15840},
  year={2020}
}