https://github.com/nikhilroxtomar/u2-net-for-image-matting-in-tensorflow
This project showcases an implementation of the U2-Net architecture for Image Matting in the TensorFlow.
https://github.com/nikhilroxtomar/u2-net-for-image-matting-in-tensorflow
image-matting portrait-matting u2-net
Last synced: 5 months ago
JSON representation
This project showcases an implementation of the U2-Net architecture for Image Matting in the TensorFlow.
- Host: GitHub
- URL: https://github.com/nikhilroxtomar/u2-net-for-image-matting-in-tensorflow
- Owner: nikhilroxtomar
- Created: 2023-05-10T15:48:57.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2023-05-10T16:18:54.000Z (about 2 years ago)
- Last Synced: 2024-11-16T06:28:00.823Z (7 months ago)
- Topics: image-matting, portrait-matting, u2-net
- Language: Python
- Homepage:
- Size: 40.5 MB
- Stars: 7
- Watchers: 1
- Forks: 7
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 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:
- [Privacy-Preserving Portrait Matting Dataset (P3M-10k)](https://drive.google.com/uc?export=download&id=1LqUU7BZeiq8I3i5KxApdOJ2haXm-cEv1)
- [P3M-10k facemask (optional)](https://drive.google.com/file/d/1I-71PbkWcivBv3ly60V0zvtYRd3ddyYs/view?usp=sharing)
## 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: