Awesome

PyTorch Semantic Segmentation

Introduction

This repository is a PyTorch implementation for semantic segmentation / scene parsing. The code is easy to use for training and testing on various datasets. The codebase mainly uses ResNet50/101/152 as backbone and can be easily adapted to other basic classification structures. Implemented networks including PSPNet and PSANet, which ranked 1st places in ImageNet Scene Parsing Challenge 2016 @ECCV16, LSUN Semantic Segmentation Challenge 2017 @CVPR17 and WAD Drivable Area Segmentation Challenge 2018 @CVPR18. Sample experimented datasets are ADE20K, PASCAL VOC 2012 and Cityscapes.

Update

2020.05.15: Branch master, use official nn.SyncBatchNorm, only multiprocessing training is supported, tested with pytorch 1.4.0.
2019.05.29: Branch 1.0.0, both multithreading training (nn.DataParallel) and multiprocessing training (nn.parallel.DistributedDataParallel) (recommended) are supported. And the later one is much faster. Use syncbn from EncNet and apex, tested with pytorch 1.0.0.

Usage

Highlight:
- Fast multiprocessing training (nn.parallel.DistributedDataParallel) with official nn.SyncBatchNorm.
- Better reimplementation results with well designed code structures.
- All initialization models, trained models and predictions are available.
Requirement:
- Hardware: 4-8 GPUs (better with >=11G GPU memory)
- Software: PyTorch>=1.1.0, Python3, tensorboardX,

Clone the repository:

git clone https://github.com/hszhao/semseg.git

Train:
- Download related datasets and symlink the paths to them as follows (you can alternatively modify the relevant paths specified in folder config):
```
cd semseg
mkdir -p dataset
ln -s /path_to_ade20k_dataset dataset/ade20k
```
- Download ImageNet pre-trained models and put them under folder initmodel for weight initialization. Remember to use the right dataset format detailed in FAQ.md.
- Specify the gpu used in config then do training:
```
sh tool/train.sh ade20k pspnet50
```
- If you are using SLURM for nodes manager, uncomment lines in train.sh and then do training:
```
sbatch tool/train.sh ade20k pspnet50
```
Test:
- Download trained segmentation models and put them under folder specified in config or modify the specified paths.
- For full testing (get listed performance):
```
sh tool/test.sh ade20k pspnet50
```
- Quick demo on one image:
```
PYTHONPATH=./ python tool/demo.py --config=config/ade20k/ade20k_pspnet50.yaml --image=figure/demo/ADE_val_00001515.jpg TEST.scales '[1.0]'
```
Visualization: tensorboardX incorporated for better visualization.
```
tensorboard --logdir=exp/ade20k
```
Other:
- Resources: GoogleDrive LINK contains shared models, visual predictions and data lists.
- Models: ImageNet pre-trained models and trained segmentation models can be accessed. Note that our ImageNet pretrained models are slightly different from original ResNet implementation in the beginning part.
- Predictions: Visual predictions of several models can be accessed.
- Datasets: attributes (names and colors) are in folder dataset and some sample lists can be accessed.
- Some FAQs: FAQ.md.
- Former video predictions: high accuracy -- PSPNet, PSANet; high efficiency -- ICNet.

Performance

Description: mIoU/mAcc/aAcc stands for mean IoU, mean accuracy of each class and all pixel accuracy respectively. ss denotes single scale testing and ms indicates multi-scale testing. Training time is measured on a sever with 8 GeForce RTX 2080 Ti. General parameters cross different datasets are listed below:

Train Parameters: sync_bn(True), scale_min(0.5), scale_max(2.0), rotate_min(-10), rotate_max(10), zoom_factor(8), ignore_label(255), aux_weight(0.4), batch_size(16), base_lr(1e-2), power(0.9), momentum(0.9), weight_decay(1e-4).
Test Parameters: ignore_label(255), scales(single: [1.0], multiple: [0.5 0.75 1.0 1.25 1.5 1.75]).

ADE20K: Train Parameters: classes(150), train_h(473/465-PSP/A), train_w(473/465-PSP/A), epochs(100). Test Parameters: classes(150), test_h(473/465-PSP/A), test_w(473/465-PSP/A), base_size(512).

Setting: train on train (20210 images) set and test on val (2000 images) set.

Network	mIoU/mAcc/aAcc(ss)	mIoU/mAcc/pAcc(ms)	Training Time
PSPNet50	0.4189/0.5227/0.8039.	0.4284/0.5266/0.8106.	14h
PSANet50	0.4229/0.5307/0.8032.	0.4305/0.5312/0.8101.	14h
PSPNet101	0.4310/0.5375/0.8107.	0.4415/0.5426/0.8172.	20h
PSANet101	0.4337/0.5385/0.8102.	0.4414/0.5392/0.8170.	20h

PSACAL VOC 2012: Train Parameters: classes(21), train_h(473/465-PSP/A), train_w(473/465-PSP/A), epochs(50). Test Parameters: classes(21), test_h(473/465-PSP/A), test_w(473/465-PSP/A), base_size(512).

Setting: train on train_aug (10582 images) set and test on val (1449 images) set.

Network	mIoU/mAcc/aAcc(ss)	mIoU/mAcc/pAcc(ms)	Training Time
PSPNet50	0.7705/0.8513/0.9489.	0.7802/0.8580/0.9513.	3.3h
PSANet50	0.7725/0.8569/0.9491.	0.7787/0.8606/0.9508.	3.3h
PSPNet101	0.7907/0.8636/0.9534.	0.7963/0.8677/0.9550.	5h
PSANet101	0.7870/0.8642/0.9528.	0.7966/0.8696/0.9549.	5h

Cityscapes: Train Parameters: classes(19), train_h(713/709-PSP/A), train_w(713/709-PSP/A), epochs(200). Test Parameters: classes(19), test_h(713/709-PSP/A), test_w(713/709-PSP/A), base_size(2048).

Setting: train on fine_train (2975 images) set and test on fine_val (500 images) set.

Network	mIoU/mAcc/aAcc(ss)	mIoU/mAcc/pAcc(ms)	Training Time
PSPNet50	0.7730/0.8431/0.9597.	0.7838/0.8486/0.9617.	7h
PSANet50	0.7745/0.8461/0.9600.	0.7818/0.8487/0.9622.	7.5h
PSPNet101	0.7863/0.8577/0.9614.	0.7929/0.8591/0.9638.	10h
PSANet101	0.7842/0.8599/0.9621.	0.7940/0.8631/0.9644.	10.5h

Citation

If you find the code or trained models useful, please consider citing:

@misc{semseg2019,
  author={Zhao, Hengshuang},
  title={semseg},
  howpublished={\url{https://github.com/hszhao/semseg}},
  year={2019}
}
@inproceedings{zhao2017pspnet,
  title={Pyramid Scene Parsing Network},
  author={Zhao, Hengshuang and Shi, Jianping and Qi, Xiaojuan and Wang, Xiaogang and Jia, Jiaya},
  booktitle={CVPR},
  year={2017}
}
@inproceedings{zhao2018psanet,
  title={{PSANet}: Point-wise Spatial Attention Network for Scene Parsing},
  author={Zhao, Hengshuang and Zhang, Yi and Liu, Shu and Shi, Jianping and Loy, Chen Change and Lin, Dahua and Jia, Jiaya},
  booktitle={ECCV},
  year={2018}
}

Question

Some FAQ.md collected. You are welcome to send pull requests or give some advices. Contact information: hengshuangzhao at gmail.com.