PyTorch implementation of MiniNet-v2

This repository contains a PyTorch implementation of MiniNet-v2 for semantic segmentation.

Setup

First, you can install the dependencies using a virtual environment. This setup has been tested with Python 3.10:

python3 -m venv env
source env/bin/activate
pip install -r requirements.txt

Dataset

We provide a script for downloading the Cityscapes dataset using the cityscapesscripts package. You will need an account in the Cityscapes website to access the data.

bash download_datasets.sh

Training

Once the dataset is downloaded, you can train a model with default settings running the training script:

python train_segmentation.py

This script by default will save checkpoints during the training in results/<date>/checkpoint_<epoch>.tar.

Inference

The inference.ipynb notebook contains an example of how to use the model for segmentation tasks. We also include a script to prepare a submission to the Cityscapes evaluation server using the test images:

python prepare_submission.py --model-path results/<date>/checkpoint_<epoch>.tar

References

• https://github.com/Shathe/MiniNet-v2
• https://github.com/endomapper/EM_Dataset-ToolSegmentation

Citation

If you find this code useful in your research, please cite the original paper:

@article{alonso2020mininet,
  title = {{MiniNet}: An Efficient Semantic Segmentation {ConvNet} for Real-Time Robotic Applications},
  author = {Alonso, I{\~n}igo and Riazuelo, Luis and Murillo, Ana C.},
  journal = {IEEE Transactions on Robotics (T-RO)},
  year = {2020},
  volume = {36},
  number = {4},
  pages = {1340--1347},
  doi = {10.1109/TRO.2020.2974099},
}