search

LinjieLyu / DPI

42 stars 2 forks source link

About roh in indoor test and outdoor test #4

Open woduck opened 2 weeks ago

woduck commented 2 weeks ago

I am currently reviewing your DPI code and the paper, and I have encountered a parameter defined as roh = 1/sigma^2. Upon examining the code, I noticed that this parameter seems to operate differently for indoor and outdoor scenarios.

Could you please provide some insight into the rationale behind this distinction? Specifically, I am curious to understand the criteria or reasoning that led to the decision to have different implementations for indoor and outdoor environments.

LinjieLyu commented 2 weeks ago

Hi there. This is because the indoor and outdoor dataset distribution is very different. The indoor Laval dataset is a real captured HDR dataset while the streetlearn outdoor dataset is an LDR dataset. Although we first transfer the outdoor dataset to an HDR dataset, we have to use distinct gamma corrections. Due to the different gamma correction parameters, the gradient from the loss function will be different so dedicated hyperparameters generate more realistic and diverse envmaps for two scenarios.