U2-Net for Image Matting in TensorFlow
Welcome to the U2-Net for Image Matting in TensorFlow repository! This project showcases an implementation of the U2-Net architecture for image matting in the TensorFlow. The code includes the training process, making use of the Privacy-Preserving Portrait Matting Dataset (P3M-10k) to achieve accurate foreground extraction from images.
Dataset
Privacy-Preserving Portrait Matting Dataset (P3M-10k) is used for training and validation process. P3M-10k contains 10421 high-resolution real-world face-blurred portrait images, along with their manually labeled alpha mattes.
Download the dataset:
Results
The sequence in the images below is as follows- Input Image, Predicted Alpha Matte and Predicted Alpha Matte applied over Input Image.
How to improve
- Train on more epochs.
- Increase the input image resolution.
- Apply data augmentation.
- Try new loss function.
Contact
For more follow me on: