search

hustvl / ViTMatte

[Information Fusion (Vol.103, Mar. '24)] Boosting Image Matting with Pretrained Plain Vision Transformers
MIT License
363 stars 37 forks source link

About gradient_loss and connectivity_error #21

Closed Wickked-swag closed 1 year ago

Wickked-swag commented 1 year ago

I found that the values of connectivity_error and gradient_loss obtained through the MATLAB file of DIM, when processing the results from inference.py, are excessively large. I wonder if these values, connectivity_error and gradient_loss, have been scaled similarly to MSE in the paper? Or is there an issue with my approach? Based on my analysis, if there are problems with the generated data, MSE and SAD should also be incorrect.

JingfengYao commented 1 year ago

The scaling of Conn and Grad has originally been done in Deep Image Matting evaluation. Its official codes did not show that clearly, but you could refer to the evaluation codes of previous famous methods IndexNet and GCAMatting.

In ViTMatte, we follow the same evaluation codes provided above. Besides, we have scaled MSE for ease of clarity.

Wickked-swag commented 1 year ago

Thank U.You solved a big doubt for me.