SimpleSeg is a lightweight, research-friendly repository for semantic segmentation from HUSTVL. It aims to provide clean baselines and utilities for training, evaluating, and deploying segmentation models in Python.
- Minimal, readable code structure for rapid experimentation
- Reproducible training and evaluation routines
- Configurable model, dataset, and augmentation pipelines
- Single- and multi-GPU training with PyTorch (if configured)
- Checkpointing, logging, and metric tracking (mIoU, pixel accuracy, etc.)
-
Download pre-trained weights: Google Drive
-
Training
python -m torch.distributed.launch --nproc_per_node ${GPUS} train.py --cfg ${CONFIG}- Evaluation
python test.py --cfg ${CONFIG} --ckpt {CKPT} TEST.BATCH_SIZE 1If you find this repository useful in your research, please cite it:
@article{gain2025,
author = {Tianheng Cheng and
Xinggang Wang and
Junchao Liao and
Wenyu Liu},
title = {Cross-layer attentive feature upsampling for low-latency semantic
segmentation},
journal = {Mach. Vis. Appl.},
volume = {36},
number = {1},
pages = {18},
year = {2025}
}- This project is maintained by HUSTVL.
- Thanks to the open-source community and prior work in semantic segmentation.