README

This is a neural network implementation based on U-NET for generating hair segmentation masks on image & video data. Built for virtual try-on of beauty products.

Details

• Architecture: U-Net network implemented with keras
• Training platform: Google colab T4 GPU
• Deployment Backend: Tensorflow JS on browser via webgl
• Input dataset: Sampled images from CelebA dataset and generated segmentation masks
• Execution speed: ~9 video frames per second

Relevant files:

• main_unet.ipynb:- training script
• demo-app:- webapp for predicting on webcam & images
• dataset:- data preparation scripts
• converter.sh:- Tensorflow to TFJS model format converter

Results on images:

prediction time: ~110ms

TODO

• Handle temporal inconsistency for videos
• Decrease prediction time