MetaSeg: MetaFormer-based Global Contexts-aware Network for Efficient Semantic Segmentation (WACV 2024)

📝Paper

Beoungwoo Kang*, Seunghun Moon*, Yubin Cho*, Hyunwoo Yu*, Suk-ju Kang^†

* Equal contribution, ^†Correspondence

Sogang University

This repository contains the official Pytorch implementation of training & evaluation code for MetaSeg.

Installation

For install and data preparation, please refer to the guidelines in MMSegmentation v0.24.1.

pip install timm
cd MetaSeg
python setup.py develop

Training

Download backbone (MSCAN-T & MSCAN-B) pretrained weights in here.

Put them in a folder pretrain/.

Example - Train MetaSeg-T on ADE20K:

CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./tools/dist_train.sh local_configs/metaseg/tiny/metaseg.tiny.512x512.ade.160k.py <GPU_NUM>

Evaluation

Download trained weights.

Example - Evaluate MetaSeg-T on ADE20K:

# Single-gpu testing
CUDA_VISIBLE_DEVICES=0 python tools/test.py local_configs/metaseg/tiny/metaseg.tiny.512x512.ade.160k.py /path/to/checkpoint_file

# Multi-gpu testing
CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./tools/dist_test.sh local_configs/metaseg/tiny/metaseg.tiny.512x512.ade.160k.py /path/to/checkpoint_file <GPU_NUM>

Citation

@inproceedings{kang2024metaseg,
  title={MetaSeg: MetaFormer-based Global Contexts-aware Network for Efficient Semantic Segmentation},
  author={Kang, Beoungwoo and Moon, Seunghun and Cho, Yubin and Yu, Hyunwoo and Kang, Suk-Ju},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={434--443},
  year={2024}
}