fcn

FCN

Fully Convolutional Networks for Semantic Segmentation

Introduction

Official Repo

Code Snippet

Abstract

Convolutional networks are powerful visual models that yield hierarchies of features. We show that convolutional networks by themselves, trained end-to-end, pixels-to-pixels, exceed the state-of-the-art in semantic segmentation. Our key insight is to build "fully convolutional" networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning. We define and detail the space of fully convolutional networks, explain their application to spatially dense prediction tasks, and draw connections to prior models. We adapt contemporary classification networks (AlexNet, the VGG net, and GoogLeNet) into fully convolutional networks and transfer their learned representations by fine-tuning to the segmentation task. We then define a novel architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentations. Our fully convolutional network achieves state-of-the-art segmentation of PASCAL VOC (20% relative improvement to 62.2% mean IU on 2012), NYUDv2, and SIFT Flow, while inference takes one third of a second for a typical image.

Citation

@article{shelhamer2017fully,
  title={Fully convolutional networks for semantic segmentation},
  author={Shelhamer, Evan and Long, Jonathan and Darrell, Trevor},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  volume={39},
  number={4},
  pages={640--651},
  year={2017},
  publisher={IEEE Trans Pattern Anal Mach Intell}
}

Results and models

Cityscapes

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
FCN	R-50-D8	512x1024	40000	5.7	4.17	72.25	73.36	config	model | log
FCN	R-101-D8	512x1024	40000	9.2	2.66	75.45	76.58	config	model | log
FCN	R-50-D8	769x769	40000	6.5	1.80	71.47	72.54	config	model | log
FCN	R-101-D8	769x769	40000	10.4	1.19	73.93	75.14	config	model | log
FCN	R-18-D8	512x1024	80000	1.7	14.65	71.11	72.91	config	model | log
FCN	R-50-D8	512x1024	80000	-		73.61	74.24	config	model | log
FCN	R-101-D8	512x1024	80000	-	-	75.13	75.94	config	model | log
FCN (FP16)	R-101-D8	512x1024	80000	5.37	8.64	76.80	-	config	model | log
FCN	R-18-D8	769x769	80000	1.9	6.40	70.80	73.16	config	model | log
FCN	R-50-D8	769x769	80000	-	-	72.64	73.32	config	model | log
FCN	R-101-D8	769x769	80000	-	-	75.52	76.61	config	model | log
FCN	R-18b-D8	512x1024	80000	1.6	16.74	70.24	72.77	config	model | log
FCN	R-50b-D8	512x1024	80000	5.6	4.20	75.65	77.59	config	model | log
FCN	R-101b-D8	512x1024	80000	9.1	2.73	77.37	78.77	config	model | log
FCN	R-18b-D8	769x769	80000	1.7	6.70	69.66	72.07	config	model | log
FCN	R-50b-D8	769x769	80000	6.3	1.82	73.83	76.60	config	model | log
FCN	R-101b-D8	769x769	80000	10.3	1.15	77.02	78.67	config	model | log
FCN (D6)	R-50-D16	512x1024	40000	3.4	10.22	77.06	78.85	config	model | log
FCN (D6)	R-50-D16	512x1024	80000	-	10.35	77.27	78.88	config	model | log
FCN (D6)	R-50-D16	769x769	40000	3.7	4.17	76.82	78.22	config	model | log
FCN (D6)	R-50-D16	769x769	80000	-	4.15	77.04	78.40	config	model | log
FCN (D6)	R-101-D16	512x1024	40000	4.5	8.04	77.36	79.18	config	model | log
FCN (D6)	R-101-D16	512x1024	80000	-	8.26	78.46	80.42	config	model | log
FCN (D6)	R-101-D16	769x769	40000	5.0	3.12	77.28	78.95	config	model | log
FCN (D6)	R-101-D16	769x769	80000	-	3.21	78.06	79.58	config	model | log
FCN (D6)	R-50b-D16	512x1024	80000	3.2	10.16	76.99	79.03	config	model | log
FCN (D6)	R-50b-D16	769x769	80000	3.6	4.17	76.86	78.52	config	model | log
FCN (D6)	R-101b-D16	512x1024	80000	4.3	8.46	77.72	79.53	config	model | log
FCN (D6)	R-101b-D16	769x769	80000	4.8	3.32	77.34	78.91	config	model | log

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
FCN	R-50-D8	512x512	80000	8.5	23.49	35.94	37.94	config	model | log
FCN	R-101-D8	512x512	80000	12	14.78	39.61	40.83	config	model | log
FCN	R-50-D8	512x512	160000	-	-	36.10	38.08	config	model | log
FCN	R-101-D8	512x512	160000	-	-	39.91	41.40	config	model | log

Pascal VOC 2012 + Aug

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
FCN	R-50-D8	512x512	20000	5.7	23.28	67.08	69.94	config	model | log
FCN	R-101-D8	512x512	20000	9.2	14.81	71.16	73.57	config	model | log
FCN	R-50-D8	512x512	40000	-	-	66.97	69.04	config	model | log
FCN	R-101-D8	512x512	40000	-	-	69.91	72.38	config	model | log

Pascal Context

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
FCN	R-101-D8	480x480	40000	-	9.93	44.43	45.63	config	model | log
FCN	R-101-D8	480x480	80000	-	-	44.13	45.26	config	model | log

Pascal Context 59

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
FCN	R-101-D8	480x480	40000	-	-	48.42	50.4	config	model | log
FCN	R-101-D8	480x480	80000	-	-	49.35	51.38	config	model | log

Note:

• FP16 means Mixed Precision (FP16) is adopted in training.
• FCN D6 means dilation rate of convolution operator in FCN is 6.