isanet

ISANet

Interlaced Sparse Self-Attention for Semantic Segmentation

Introduction

Official Repo

Code Snippet

Abstract

In this paper, we present a so-called interlaced sparse self-attention approach to improve the efficiency of the \emph{self-attention} mechanism for semantic segmentation. The main idea is that we factorize the dense affinity matrix as the product of two sparse affinity matrices. There are two successive attention modules each estimating a sparse affinity matrix. The first attention module is used to estimate the affinities within a subset of positions that have long spatial interval distances and the second attention module is used to estimate the affinities within a subset of positions that have short spatial interval distances. These two attention modules are designed so that each position is able to receive the information from all the other positions. In contrast to the original self-attention module, our approach decreases the computation and memory complexity substantially especially when processing high-resolution feature maps. We empirically verify the effectiveness of our approach on six challenging semantic segmentation benchmarks.

Results and models

Cityscapes

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
ISANet	R-50-D8	512x1024	40000	5.869	2.91	V100	78.49	79.44	config	model | log
ISANet	R-50-D8	512x1024	80000	5.869	2.91	V100	78.68	80.25	config	model | log
ISANet	R-50-D8	769x769	40000	6.759	1.54	V100	78.70	80.28	config	model | log
ISANet	R-50-D8	769x769	80000	6.759	1.54	V100	79.29	80.53	config	model | log
ISANet	R-101-D8	512x1024	40000	9.425	2.35	V100	79.58	81.05	config	model | log
ISANet	R-101-D8	512x1024	80000	9.425	2.35	V100	80.32	81.58	config	model | log
ISANet	R-101-D8	769x769	40000	10.815	0.92	V100	79.68	80.95	config	model | log
ISANet	R-101-D8	769x769	80000	10.815	0.92	V100	80.61	81.59	config	model | log

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
ISANet	R-50-D8	512x512	80000	9.0	22.55	V100	41.12	42.35	config	model | log
ISANet	R-50-D8	512x512	160000	9.0	22.55	V100	42.59	43.07	config	model | log
ISANet	R-101-D8	512x512	80000	12.562	10.56	V100	43.51	44.38	config	model | log
ISANet	R-101-D8	512x512	160000	12.562	10.56	V100	43.80	45.4	config	model | log

Pascal VOC 2012 + Aug

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
ISANet	R-50-D8	512x512	20000	5.9	23.08	V100	76.78	77.79	config	model | log
ISANet	R-50-D8	512x512	40000	5.9	23.08	V100	76.20	77.22	config	model | log
ISANet	R-101-D8	512x512	20000	9.465	7.42	V100	78.46	79.16	config	model | log
ISANet	R-101-D8	512x512	40000	9.465	7.42	V100	78.12	79.04	config	model | log

Citation

@article{huang2019isa,
  title={Interlaced Sparse Self-Attention for Semantic Segmentation},
  author={Huang, Lang and Yuan, Yuhui and Guo, Jianyuan and Zhang, Chao and Chen, Xilin and Wang, Jingdong},
  journal={arXiv preprint arXiv:1907.12273},
  year={2019}
}

The technical report above is also presented at:

@article{yuan2021ocnet,
  title={OCNet: Object Context for Semantic Segmentation},
  author={Yuan, Yuhui and Huang, Lang and Guo, Jianyuan and Zhang, Chao and Chen, Xilin and Wang, Jingdong},
  journal={International Journal of Computer Vision},
  pages={1--24},
  year={2021},
  publisher={Springer}
}