FDLNet

This is the official implementation of paper "FDLNet: Boosting Real-time Semantic Segmentation by Image-size Convolution via Frequency Domain Learning" (PyTorch | MMSegmentation) (ICRA 2023).

Getting started

This repo is based on the MMSegmentation project. To use this repo, please:

1. Install MMSegmentation following the official instructions:

# a. Clone the official repo
git clone -b 0.x https://github.com/open-mmlab/mmsegmentation.git

# b. Install MMCV using MIM
pip install -U openmim
mim install mmcv-full

# c. Install MMSegmentation
cd mmsegmentation
pip install -v -e .

2. Merge this repo into the official MMSegmention:

# Merge
git clone https://github.com/qyan0131/FDLNet.git
cd FDLNet
cp -r * mmsegmentation/

Then register FDLNet by:

a. Please add from .fdlnet import FDLNet to mmseg/models/backbones/__init__.py.

b. Append 'FDLNet' at the end of entrance __all__=[...].

ALTERNATIVELY, this repo provides a patch for git apply:

cd /path/to/mmsegmentation
git reset --hard 0beaf6
git apply fdlnet.patch

Train

This repo follows the official training pipeline of MMSegmentation:

tools/dist_train.sh configs/fdlnet/fdlnet18_b6x2_cityscapes_512x1024.py 2

Citation

If you find this work/repo helpful, please consider citing:

@INPROCEEDINGS{yan2023fdlnet,
  author    = {Yan, Qingqing and Li, Shu and Liu, Chengju and Liu, Ming and Chen, Qijun},
  booktitle = {2023 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title     = {FDLNet: Boosting Real-time Semantic Segmentation by Image-size Convolution via Frequency Domain Learning}, 
  year      = {2023},
  pages     = {8155-8162},
  doi       = {10.1109/ICRA48891.2023.10161421}
  }

Acknowledgement

This code is built using MMSegmentation repo.