README.md

Boosting Night-time Scene Parsing with
Learnable Frequency

This repo is the official implementation of "Boosting Night-time Scene Parsing with Learnable Frequency (IEEE TIP 2023) ".

Installation

Our work is based on " awesome-semantic-segmentation-pytorch", please follow their README.md for installation.

Data Preparation

"NightCity"

"NightCity+" (Only reannotated val set)

"BDD100K-night" (Only images, please download the labels from here with permission)

Train

cd scripts
python train_edge.py --model fdlnet --backbone resnet50 --dataset night --aux

Test

cd scripts
python eval.py --model fdlnet --backbone resnet101 --dataset night --aux

Results and Models

Dataset	mIoU	w/ ms	Model
NightCity	54.60	55.42	FDLNet (DeeplabV3+)
NightCity+	56.20	56.79	~

Citation

If you find this repo useful for your research, please consider citing our paper:

@ARTICLE{10105211,
  author={Xie, Zhifeng and Wang, Sen and Xu, Ke and Zhang, Zhizhong and Tan, Xin and Xie, Yuan and Ma, Lizhuang},
  journal={IEEE Transactions on Image Processing}, 
  title={Boosting Night-Time Scene Parsing With Learnable Frequency}, 
  year={2023},
  volume={32},
  pages={2386-2398},
  doi={10.1109/TIP.2023.3267044}}