Skip to content

Latest commit

 

History

History
21 lines (12 loc) · 739 Bytes

README.md

File metadata and controls

21 lines (12 loc) · 739 Bytes
Raw

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