BiAlignNet

This repository contains the source code of BiAlignNet (Bidirectional Alignment Network) from the paper FAST AND ACCURATE SCENE PARSING VIA BI-DIRECTION ALIGNMENT NETWORKS, ICIP 2021.

Instructions

Please follow the instructions to run the code.

BialignNet-Dfnet2: mIoU 78.7

Prerequisite

Read the DATASET.md to prepare the dataset you want to use (in this repo, we use cityscapes)
Set the dataset path in config.py
In this directory, mkdir pretrained_models
Download the pretrained dfnet1 and dfnet2 and save them into pretrained_models directory

Training

sh scripts/train/train_cityscapes_bialign_dfnet2.sh

Evaluation in validation set

sh scripts/evaluate_val/eval_cityscapes_bialign_dfnet2.sh /path/to/model /path/to/where/you/want/to/save/results

Submitting the test set result

sh scripts/submit_test/submit_cityscapes_bialign.sh /path/to/model /path/to/where/you/want/to/save/results

Citation

@inproceedings{wu2021fast,
  title={Fast and Accurate Scene Parsing via Bi-Direction Alignment Networks},
  author={Wu, Yanran and Li, Xiangtai and Shi, Chen and Tong, Yunhai and Hua, Yang and Song, Tao and Ma, Ruhui and Guan, Haibing},
  booktitle={2021 IEEE International Conference on Image Processing (ICIP)},
  pages={2508--2512},
  year={2021},
  organization={IEEE}
}

Acknowledgement

This repo is based on Semantic Segmentation from SFSegNets.

jojacola/BiAlignNet