Implementation of the paper "SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder" for portrait segmentation in PyTorch including pretrained weights and training script reaching 95.25 IoU on the EG1800 dataset (trained with same schedule and data as original authors).
- place data in "datasets" folder (data available through https://github.com/HYOJINPARK/ExtPortraitSeg)
ls -l datasets
EG1800
Nukki
--\
baidu_V1
baidu_V2
Install requirements
pip install -r requirements.txt
Run training
python train.py [--skip-encoder] [--use-cuda]
# Todo
Fix issue with different input image size than 244