MMSegmentation - SFNet

基于 mmsegmentation 的 SFNet 复现，提交于 OpenMMLab 算法生态挑战赛。

原始论文：Semantic Flow for Fast and Accurate Scene Parsing

阅读其他语言的版本：English | 简体中文

目录

• 环境要求
• 目录结构
• 用法
  • 训练
  • 测试
• 结果
• 引用

环境要求

• mim == v0.1.0
• mmcv-full == v1.3.8
• mmsegmentation == v0.15.0

$ pip install openmim
$ pip install mmcv-full
$ mim install mmsegmentation

目录结构

SFNet
  |---- configs
  |       |---- _base_
  |       |       |---- datasets
  |       |       |        |---- cityscapes.py
  |       |       |---- models
  |       |       |        |---- sfnet_r18-d8_dsn.py
  |       |       |        |---- sfnet_r18-d8.py
  |       |       |        |---- sfnet_r50-d8.py
  |       |       |        |---- sfnet_r101-d8.py
  |       |       |---- schedules
  |       |       |        |---- schedule_50k.py
  |       |       |---- default_runtime.py
  |       |---- sfnet
  |               |---- sfnet_r18-d8_512x1024_50k_cityscapes.py
  |               |---- sfnet_r18-d8_dsn_512x1024_50k_cityscapes.py
  |               |---- sfnet_r50-d8_512x1024_50k_cityscapes.py
  |               |---- sfnet_r101-d8_512x1024_50k_cityscapes.py
  |---- mmseg
  |       |---- models
  |               |---- backbones
  |               |       |---- resnet.py
  |               |---- necks
  |                       |---- sf_neck.py
  |---- .gitignore
  |---- README_en-US.md
  |---- README.md
  |---- requirements.txt
  |---- setup.cfg

用法

训练

$ PYTHONPATH='.':$PYTHONPATH mim train mmseg configs/sfnet/sfnet_r18-d8_512x1024_50k_cityscapes.py \
    --work-dir ./work_dirs/sfnet_r18-d8_512x1024_50k_cityscapes/ \
    --gpus 1

测试

由于 Cityscapes 官方不提供测试集 Ground Truth，所以首先需要为测试集中的图像生成分割结果：

$ PYTHONPATH='.':$PYTHONPATH mim test mmseg configs/sfnet/sfnet_r18-d8_512x1024_50k_cityscapes.py \
    --checkpoint ./work_dirs/sfnet_r18-d8_512x1024_50k_cityscapes/epoch_300.pth \
    --format-only \
    --eval-options "imgfile_prefix=./results"

然后将所有分割结果打包压缩：

$ zip -r results.zip results/

最后手动将该压缩包上传到 Cityscapes 官网的提交入口即可。

结果

比赛要求在 cityscapes 上达到超过 77% 的 mIoU，复现得到的三个模型均达到了超过 78% 的 mIoU，满足精度要求。

整体结果：

模型	aAcc	mAcc	mIoU	FLOPs	Params	链接
SFNet(ResNet-18)	96.0	84.47	78.03	243.32G	12.87M	sfnet-resnet18-dsn.pth
SFNet(ResNet-18) +dsn	95.97	84.89	78.31	243.32G	13.32M	sfnet-resnet18.pth
SFNet(ResNet-50)	96.42	86.23	80.17	664.07G	31.33M	sfnet-resnet50.pth
SFNet(ResNet-101)	96.47	87.38	81.01	819.46G	50.32M	sfnet-resnet101.pth

在验证集各个类别上的 mIoU：

类别	SFNet(ResNet-18)	SFNet(ResNet-50)	SFNet(ResNet-101)
road	97.94	98.29	98.18
sidewalk	83.78	85.88	85.39
building	92.29	93.08	93.42
wall	58.26	60.17	61.86
fence	61.92	63.84	65.67
pole	62.6	67.43	68.5
traffic light	70.05	73.13	73.6
traffic sign	77.61	80.73	81.4
vegetation	92.19	92.86	92.85
terrain	63.74	65.15	65.55
sky	94.52	95.1	95.12
person	81.75	83.03	84.36
rider	62.46	65.28	69.83
car	94.97	95.58	95.77
truck	83.09	81.0	85.65
bus	88.65	91.15	90.69
train	81.89	84.39	84.81
motorcycle	62.8	67.72	67.23
bicycle	77.32	79.47	79.31

引用

@misc{mmcv,
    title={{MMCV: OpenMMLab} Computer Vision Foundation},
    author={MMCV Contributors},
    howpublished = {\url{https://github.com/open-mmlab/mmcv}},
    year={2018}
}
@misc{mmseg2020,
    title={{MMSegmentation}: OpenMMLab Semantic Segmentation Toolbox and Benchmark},
    author={MMSegmentation Contributors},
    howpublished = {\url{https://github.com/open-mmlab/mmsegmentation}},
    year={2020}
}