search

Kai-46 / IRON

Inverse rendering by optimizing neural SDF and materials from photometric images
BSD 2-Clause "Simplified" License
299 stars 24 forks source link

The question about the edge gradient problem existed in IDR #22

Open pean1128 opened 1 year ago

pean1128 commented 1 year ago

Hi, I have read this paper and some questions appear. I agree with this paragraph that "For this reason, these methods compute biased gradients w.r.t. the weights of the neural SDF that move surface points along camera rays”. It is easy to understand due to the loss function mentioned in IDR. However, you also display Figure.4 to demonstrate your efficiency which immediately confuses me. Because IDR adopts mask loss to train so that the SDF value will be optimized to fit silhouette supervision and won't lead to similar degradation. Can you give me some inspiration or advice?

Jarvisss commented 1 year ago

I have the same question. @Kai-46 Could you give some insight on this question?

Hi, I have read this paper and some questions appear. I agree with this paragraph that "For this reason, these methods compute biased gradients w.r.t. the weights of the neural SDF that move surface points along camera rays”. It is easy to understand due to the loss function mentioned in IDR. However, you also display Figure.4 to demonstrate your efficiency which immediately confuses me. Because IDR adopts mask loss to train so that the SDF value will be optimized to fit silhouette supervision and won't lead to similar degradation. Can you give me some inspiration or advice?

Kai-46 commented 1 year ago

We assume the inputs do not include object segmentation masks; hence the mask loss in IDR is not applicable.