Eyes Lips Segmentation using DeepLabv3 and UNet

PyTorch code for binary segmentation on the CelebAMask-HQ dataset in an intuitive and well-documented Jupyter Notebook you can follow along as a tutorial. We train two models - a UNet architecture from scratch and a pretrained DeepLabv3 with a ResNet50 backbone.

Preparing the data

Run the script make_masks.py after downloading and unzipping the CelebAMask-HQ dataset to generate eyes and lips segmented masks. Or just obtain the pre-made masks from here.

UNet

• Annoted architecture :

• Notebook :

• Pretrained weights after running 20 epochs on 1000 samples :

DeepLabv3

• Notebook :
• Pretrained weights after running 15 epochs on 30000 samples :

Results