gcnet

GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond

Introduction

Official Repo

Code Snippet

Abstract

The Non-Local Network (NLNet) presents a pioneering approach for capturing long-range dependencies, via aggregating query-specific global context to each query position. However, through a rigorous empirical analysis, we have found that the global contexts modeled by non-local network are almost the same for different query positions within an image. In this paper, we take advantage of this finding to create a simplified network based on a query-independent formulation, which maintains the accuracy of NLNet but with significantly less computation. We further observe that this simplified design shares similar structure with Squeeze-Excitation Network (SENet). Hence we unify them into a three-step general framework for global context modeling. Within the general framework, we design a better instantiation, called the global context (GC) block, which is lightweight and can effectively model the global context. The lightweight property allows us to apply it for multiple layers in a backbone network to construct a global context network (GCNet), which generally outperforms both simplified NLNet and SENet on major benchmarks for various recognition tasks. The code and configurations are released at this https URL.

GCNet (ICCVW'2019/TPAMI'2020)

@inproceedings{cao2019gcnet,
  title={Gcnet: Non-local networks meet squeeze-excitation networks and beyond},
  author={Cao, Yue and Xu, Jiarui and Lin, Stephen and Wei, Fangyun and Hu, Han},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision Workshops},
  pages={0--0},
  year={2019}
}

Results and models

Cityscapes

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
GCNet	R-50-D8	512x1024	40000	5.8	3.93	77.69	78.56	config	model | log
GCNet	R-101-D8	512x1024	40000	9.2	2.61	78.28	79.34	config	model | log
GCNet	R-50-D8	769x769	40000	6.5	1.67	78.12	80.09	config	model | log
GCNet	R-101-D8	769x769	40000	10.5	1.13	78.95	80.71	config	model | log
GCNet	R-50-D8	512x1024	80000	-	-	78.48	80.01	config	model | log
GCNet	R-101-D8	512x1024	80000	-	-	79.03	79.84	config	model | log
GCNet	R-50-D8	769x769	80000	-	-	78.68	80.66	config	model | log
GCNet	R-101-D8	769x769	80000	-	-	79.18	80.71	config	model | log

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
GCNet	R-50-D8	512x512	80000	8.5	23.38	41.47	42.85	config	model | log
GCNet	R-101-D8	512x512	80000	12	15.20	42.82	44.54	config	model | log
GCNet	R-50-D8	512x512	160000	-	-	42.37	43.52	config	model | log
GCNet	R-101-D8	512x512	160000	-	-	43.69	45.21	config	model | log

Pascal VOC 2012 + Aug

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
GCNet	R-50-D8	512x512	20000	5.8	23.35	76.42	77.51	config	model | log
GCNet	R-101-D8	512x512	20000	9.2	14.80	77.41	78.56	config	model | log
GCNet	R-50-D8	512x512	40000	-	-	76.24	77.63	config	model | log
GCNet	R-101-D8	512x512	40000	-	-	77.84	78.59	config	model | log