Inofficial PyTorch implementation of Box2Pix: Single-Shot Instance Segmentation by Assigning Pixels to Object Boxes (Uhrig et al., 2018).
This is needed to get the project in a state where it can be trained:
- mAP metric
Instance segmentation can be added later as it's just a post processing step.
- Install PyTorch (pytorch.org)
pip install -r requirements.txt(Currently requires torchvision master)- Download the Cityscapes dataset
Train model:
python train.py --dataset-dir 'data/cityscapes'