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.
Please follow the instructions to run the code.
BialignNet-Dfnet2: mIoU 78.7
- 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
sh scripts/train/train_cityscapes_bialign_dfnet2.sh
sh scripts/evaluate_val/eval_cityscapes_bialign_dfnet2.sh /path/to/model /path/to/where/you/want/to/save/results
sh scripts/submit_test/submit_cityscapes_bialign.sh /path/to/model /path/to/where/you/want/to/save/results
@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}
}
This repo is based on Semantic Segmentation from SFSegNets.