bisenetv1

BiSeNetV1

BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation

Introduction

Official Repo

Code Snippet

Abstract

Semantic segmentation requires both rich spatial information and sizeable receptive field. However, modern approaches usually compromise spatial resolution to achieve real-time inference speed, which leads to poor performance. In this paper, we address this dilemma with a novel Bilateral Segmentation Network (BiSeNet). We first design a Spatial Path with a small stride to preserve the spatial information and generate high-resolution features. Meanwhile, a Context Path with a fast downsampling strategy is employed to obtain sufficient receptive field. On top of the two paths, we introduce a new Feature Fusion Module to combine features efficiently. The proposed architecture makes a right balance between the speed and segmentation performance on Cityscapes, CamVid, and COCO-Stuff datasets. Specifically, for a 2048x1024 input, we achieve 68.4% Mean IOU on the Cityscapes test dataset with speed of 105 FPS on one NVIDIA Titan XP card, which is significantly faster than the existing methods with comparable performance.

Results and models

Cityscapes

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
BiSeNetV1	R-18-D32 (No Pretrain)	1024x1024	160000	5.69	31.77	V100	74.44	77.05	config	model | log
BiSeNetV1	R-18-D32	1024x1024	160000	5.69	31.77	V100	74.37	76.91	config	model | log
BiSeNetV1	R-18-D32 (4x8)	1024x1024	160000	11.17	31.77	V100	75.16	77.24	config	model | log
BiSeNetV1	R-50-D32 (No Pretrain)	1024x1024	160000	15.39	7.71	V100	76.92	78.87	config	model | log
BiSeNetV1	R-50-D32	1024x1024	160000	15.39	7.71	V100	77.68	79.57	config	model | log

COCO-Stuff 164k

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
BiSeNetV1	R-18-D32 (No Pretrain)	512x512	160000	-	-	V100	25.45	26.15	config	model | log
BiSeNetV1	R-18-D32	512x512	160000	6.33	74.24	V100	28.55	29.26	config	model | log
BiSeNetV1	R-50-D32 (No Pretrain)	512x512	160000	-	-	V100	29.82	30.33	config	model | log
BiSeNetV1	R-50-D32	512x512	160000	9.28	32.60	V100	34.88	35.37	config	model | log
BiSeNetV1	R-101-D32 (No Pretrain)	512x512	160000	-	-	V100	31.14	31.76	config	model | log
BiSeNetV1	R-101-D32	512x512	160000	10.36	25.25	V100	37.38	37.99	config	model | log

Note:

• 4x8: Using 4 GPUs with 8 samples per GPU in training.
• For BiSeNetV1 on Cityscapes dataset, default setting is 4 GPUs with 4 samples per GPU in training.
• No Pretrain means the model is trained from scratch.

Citation

@inproceedings{yu2018bisenet,
  title={Bisenet: Bilateral segmentation network for real-time semantic segmentation},
  author={Yu, Changqian and Wang, Jingbo and Peng, Chao and Gao, Changxin and Yu, Gang and Sang, Nong},
  booktitle={Proceedings of the European conference on computer vision (ECCV)},
  pages={325--341},
  year={2018}
}