yucornetto / MGMatting

This repository includes the official project of Mask Guided (MG) Matting, presented in our paper: Mask Guided Matting via Progressive Refinement Network
Other
332 stars 48 forks source link

What's the difference between the DIM and RWP ? #18

Open dreamer121121 opened 3 years ago

dreamer121121 commented 3 years ago

I train the network using the DIM config, but got very poor results, This is my train loss curve, and should i use the RWP config ?

![Uploading 截屏2021-07-26 下午7.07.31.png…]()

jsjxyjl commented 3 years ago

hi ,Do you have any other problems with your training?

yucornetto commented 3 years ago

Hi, if you are training the model for DIM benchmark, please use DIM config, if for real-world portrait images, please use RWP config and the data preparation.

dreamer121121 commented 3 years ago

why comp weight set to 0 in config file, why not use comp loss to train network

kartynnik commented 3 years ago

why comp weight set to 0 in config file, why not use comp loss to train network

As mentioned in the paper, during RWP training the composited input is heavily augmented (e.g. with noise), whereas in the compositional loss it would be compared against a blend of clean, non-augmented source background and foreground, which makes no sense.