deeplabv3

Rethinking atrous convolution for semantic image segmentation

Introduction

@article{chen2017rethinking,
  title={Rethinking atrous convolution for semantic image segmentation},
  author={Chen, Liang-Chieh and Papandreou, George and Schroff, Florian and Adam, Hartwig},
  journal={arXiv preprint arXiv:1706.05587},
  year={2017}
}

Results and models

Note: D-8 here corresponding to the output stride 8 setting for DeepLab series.

Cityscapes

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	download
DeepLabV3	R-50-D8	512x1024	40000	6.1	2.57	79.09	80.45	model | log
DeepLabV3	R-101-D8	512x1024	40000	9.6	1.92	77.12	79.61	model | log
DeepLabV3	R-50-D8	769x769	40000	6.9	1.11	78.58	79.89	model | log
DeepLabV3	R-101-D8	769x769	40000	10.9	0.83	79.27	80.11	model | log
DeepLabV3	R-50-D8	512x1024	80000	-	-	79.32	80.57	model | log
DeepLabV3	R-101-D8	512x1024	80000	-	-	80.20	81.21	model | log
DeepLabV3	R-50-D8	769x769	80000	-	-	79.89	81.06	model | log
DeepLabV3	R-101-D8	769x769	80000	-	-	79.67	80.81	model | log

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	download
DeepLabV3	R-50-D8	512x512	80000	8.9	14.76	42.42	43.28	model | log
DeepLabV3	R-101-D8	512x512	80000	12.4	10.14	44.08	45.19	model | log
DeepLabV3	R-50-D8	512x512	160000	-	-	42.66	44.09	model | log
DeepLabV3	R-101-D8	512x512	160000	-	-	45.00	46.66	model | log

Pascal VOC 2012 + Aug

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	download
DeepLabV3	R-50-D8	512x512	20000	6.1	13.88	76.17	77.42	model | log
DeepLabV3	R-101-D8	512x512	20000	9.6	9.81	78.70	79.95	model | log
DeepLabV3	R-50-D8	512x512	40000	-	-	77.68	78.78	model | log
DeepLabV3	R-101-D8	512x512	40000	-	-	77.92	79.18	model | log