Bidirectional Pyramid Networks for Semantic Segmentation

by Nie, D., Xue, J. and Ren, X., details can be found here

Introduction

This repository is build for the proposed Bidirectional Pyramid Networks (BPNet), which contains full training and testing code on several segmentation datasets.

Usage

Requirement:
- Hardware: tested with RTX 2080 TI (11G).
- Software: tested with PyTorch 1.2.0, Python3.7, CUDA 10.0, tensorboardX, Ninja, tqdm, Easydict
- Anaconda is strongly recommended
Clone the repository:

 
   git clone https://github.com/ginobilinie/BPNet.git

How to Train
- create the config file of dataset:train.txt, val.txt, test.txt
file structure：(split with tab)
```
  path-of-the-image   path-of-the-groundtruth
```
- modify the config.py according to your requirements
- train a network:
```
  export NGPUS=8
  python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py  
```

How to Test

inference

  python eval -e epoch_idx -d device_idx [--verbose ] [--show_image] [--save_path Pred_Save_Path]

Visualization
Other Resources

Resources: GoogleDrive LINK contains pretrained models and some share models.

Performance

Train Parameters for BPNet_S4: steps(1100), epochs(80), base_lr(0.1), lr_scheduler(poly), momentum(0.9), weight_decay(1e-4).
Train Parameters for BPNet_S3: steps(1100), epochs(80), base_lr(0.1), lr_scheduler(poly), momentum(0.9), weight_decay(1e-4).

Cite

@inproceedings{nie2020bidirectional,
  title={Bidirectional Pyramid Networks for Semantic Segmentation},
  author={Nie, Dong and Xue, Jia and Ren, Xiaofeng},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  year={2020}
}

Thanks

Our work uses part of codes from https://github.com/ycszen/TorchSeg and https://github.com/MendelXu/ANN. Thanks for the great work!