ICNet
ICNet for Real-time Semantic Segmentation on High-resolution Images
Introduction
Official Repo
Code Snippet
Abstract
We focus on the challenging task of real-time semantic segmentation in this paper. It finds many practical applications and yet is with fundamental difficulty of reducing a large portion of computation for pixel-wise label inference. We propose an image cascade network (ICNet) that incorporates multi-resolution branches under proper label guidance to address this challenge. We provide in-depth analysis of our framework and introduce the cascade feature fusion unit to quickly achieve high-quality segmentation. Our system yields real-time inference on a single GPU card with decent quality results evaluated on challenging datasets like Cityscapes, CamVid and COCO-Stuff.
Citation
@inproceedings{zhao2018icnet,
title={Icnet for real-time semantic segmentation on high-resolution images},
author={Zhao, Hengshuang and Qi, Xiaojuan and Shen, Xiaoyong and Shi, Jianping and Jia, Jiaya},
booktitle={Proceedings of the European conference on computer vision (ECCV)},
pages={405--420},
year={2018}
}
Results and models
Cityscapes
Method |
Backbone |
Crop Size |
Lr schd |
Mem (GB) |
Inf time (fps) |
mIoU |
mIoU(ms+flip) |
config |
download |
ICNet |
R-18-D8 |
832x832 |
80000 |
1.70 |
27.12 |
68.14 |
70.16 |
config |
model | log |
ICNet |
R-18-D8 |
832x832 |
160000 |
- |
- |
71.64 |
74.18 |
config |
model | log |
ICNet (in1k-pre) |
R-18-D8 |
832x832 |
80000 |
- |
- |
72.51 |
74.78 |
config |
model | log |
ICNet (in1k-pre) |
R-18-D8 |
832x832 |
160000 |
- |
- |
74.43 |
76.72 |
config |
model | log |
ICNet |
R-50-D8 |
832x832 |
80000 |
2.53 |
20.08 |
68.91 |
69.72 |
config |
model | log |
ICNet |
R-50-D8 |
832x832 |
160000 |
- |
- |
73.82 |
75.67 |
config |
model | log |
ICNet (in1k-pre) |
R-50-D8 |
832x832 |
80000 |
- |
- |
74.58 |
76.41 |
config |
model | log |
ICNet (in1k-pre) |
R-50-D8 |
832x832 |
160000 |
- |
- |
76.29 |
78.09 |
config |
model | log |
ICNet |
R-101-D8 |
832x832 |
80000 |
3.08 |
16.95 |
70.28 |
71.95 |
config |
model | log |
ICNet |
R-101-D8 |
832x832 |
160000 |
- |
- |
73.80 |
76.10 |
config |
model | log |
ICNet (in1k-pre) |
R-101-D8 |
832x832 |
80000 |
- |
- |
75.57 |
77.86 |
config |
model | log |
ICNet (in1k-pre) |
R-101-D8 |
832x832 |
160000 |
- |
- |
76.15 |
77.98 |
config |
model | log |
Note: in1k-pre
means pretrained model is used.