High_resolution_matting_PaddlePaddle

The PaddlePaddle re-implementation of Real-Time High-Resolution Background Matting.

Team: 光焰合体技

Paper Name: Real-Time High-Resolution Background Matting

Target: a real-time, high-resolution background replacement technique

Paper Official Code: https://github.com/PeterL1n/BackgroundMattingV2

Dataset: https://grail.cs.washington.edu/projects/background-matting-v2/#/datasets

Download address: https://drive.google.com/file/d/1KpHKYW986Dax9-ZIM7I-HyBoWVcLPuaQ/view

Stage1: Run the official Code

I use the Colab from Google to run the code:

https://colab.research.google.com/drive/1r2sHzcMh5S9ABTeFqIVGQKJYMe9Rz8El

Stage2: Dataset

VideoMatte240K_JPEG_HD

https://drive.google.com/file/d/1IUp_301x8BnPjE81QBzyLASn3ZSosUF6/view

VideoMatte240K_JPEG_SD

https://drive.google.com/u/0/uc?id=1-S4F-rB75E8I7YUpHfu3itIl1knFhhFF&export=download

PhotoMatte85 https://drive.google.com/file/d/1KpHKYW986Dax9-ZIM7I-HyBoWVcLPuaQ/view?usp=sharing

stage3 Network Transfer from Pytorch to PaddlePaddle

Network Transfer is finished.

Total Training Data Pairs: 237989

Tranferring the OPs in Pytorch to Native C++ implementations.