Self-attention on Multi-shifted Windows for Scene Segmentation

Introduction

This is the PyTorch implementation of our [Self-attention on Multi-shifted Windows for Scene Segmentation] by Litao Yu, Zhibin Li and Jian Zhang

Use of code

This implementation is built based on mmsegmentation (https://mmsegmentation.readthedocs.io/). Please follow their instructions to install the necessary modules for running the programs.

Results and models

Pascal VOC 2012 + Aug

Method	Backbone	Crop Size	Lr schd	mIoU	mIoU(ms+flip)	config	download
MSwin-P	Swin-S	512x512	80000	81.58	82.95	config	model
MSwin-S	Swin-S	512x512	80000	81.97	82.74	config	model
MSwin-C	Swin-S	512x512	80000	81.40	82.70	config	model
MSwin-P	Swin-B	512x512	80000	82.85	83.82	config	model
MSwin-S	Swin-B	512x512	80000	82.86	84.27	config	model
MSwin-C	Swin-B	512x512	80000	83.50	84.47	config	model

Pascal COCOStuff-10K

Method	Backbone	Crop Size	Lr schd	mIoU	mIoU(ms+flip)	config	download
MSwin-P	Swin-S	480x480	40000	39.8	41.1	config	model
MSwin-S	Swin-S	480x480	40000	40.2	42.1	config	model
MSwin-C	Swin-S	480x480	40000	39.6	41.6	config	model
MSwin-P	Swin-B	480x480	40000	41.3	42.7	config	model
MSwin-S	Swin-B	480x480	40000	41.1	42.4	config	model
MSwin-C	Swin-B	480x480	40000	41.0	42.8	config	model

ADE20K

Method	Backbone	Crop Size	Lr schd	mIoU	mIoU(ms+flip)	config	download
MSwin-P	Swin-S	512x512	160000	47.11	48.55	config	model
MSwin-S	Swin-S	512x512	160000	47.52	48.56	config	model
MSwin-C	Swin-S	512x512	160000	46.26	48.12	config	model
MSwin-P	Swin-B	512x512	160000	48.38	50.29	config	model
MSwin-S	Swin-B	512x512	160000	48.54	50.26	config	model
MSwin-C	Swin-B	512x512	160000	48.70	50.13	config	model

Cityscapes

Method	Backbone	Crop Size	Lr schd	mIoU	mIoU(ms+flip)	config	download
MSwin-P	Swin-B	512x1024	80000	81.06	82.10	config	model
MSwin-S	Swin-B	512x1024	80000	80.87	82.39	config	model
MSwin-C	Swin-B	512x1024	80000	80.78	82.04	config	model

yutao1008/MSwin