search

husheng12345 / Unofficial-NeRF-Supervised-Deep-Stereo

An unofficial implementation of NeRF-Supervised Deep Stereo (CVPR 2023)
MIT License
17 stars 1 forks source link

AO image #3

Closed AI-student-wl closed 11 months ago

AI-student-wl commented 11 months ago

Thank you for your selfless dedication. I have a question about the data and hope to get your help. In the official open-source data, there is an AO (Ambient Occlusion) image. I'm not sure if you have rendered it using the NeRF model. The AO I rendered is all 1 or 0.99999. If you could give me some guidance, I would be extremely grateful. Once again, thank you for your open-source work and I wish you a pleasant life.

husheng12345 commented 11 months ago

Could you please provide more detailed information, such as which specific NeRF model you are using?

AI-student-wl commented 11 months ago

I am using nerfstudio. Below is my code for generating depths image

AI-student-wl commented 11 months ago

Sorry, according to my understanding of the paper, weights_ Sum is AO, but after saving it, it's all 1, which confuses me

husheng12345 commented 11 months ago

Firstly, I must mention that since I have not used nerfstudio myself, I am unable to fully answer your question. However, it is common for NeRF implementations to treat the furthest radiance as an opaque wall to ensure that the AO strictly equals 1. Therefore, you would simply remove that "wall" in your calculations to correctly obtain AO values less than 1.

AI-student-wl commented 11 months ago

Thank you for your prompt. I will follow this approach to check the code. Thank you for your help