search

rover-xingyu / Ha-NeRF

[CVPR 2022] Ha-NeRF😆: Hallucinated Neural Radiance Fields in the Wild
148 stars 13 forks source link

About view consistency loss #4

Closed y6216886 closed 2 years ago

y6216886 commented 2 years ago

Thanks for your inspiring work.

I have a question about figure 3, in which you render I^{r} {i} based on camera ray and appearance feature of I {i}.

I believe that the image I^{i} for generating the appearance feature, and the ground truth image for supervising rendered image I^{r} _{i} should be from the same image domain in which images share the same appearance feature.

However, how to identify the image domain in the given datasets?

I checked the brandenburg_gate dataset, the informations are image name, id, split, and name of dataset. I wonder which information tells the set of images is from the same image domain?

y6216886 commented 2 years ago

Ps: From Fig.3, I {i} and I^{r} {i} are images from different camera pose, that is why I ask this question.

y6216886 commented 2 years ago

image

rover-xingyu commented 2 years ago

Thanks for your interest! Maybe you have some misconceptions about the view-consistent loss, there is no ground truth image for supervising the rendered image I^{r} {i}, and that's why we built the view-consistent loss to supervise the hallucinated image I^{r} {i}.

xuxumiao777 commented 4 months ago

Where is the view consistent loss in your code?

rover-xingyu commented 4 months ago

https://github.com/rover-xingyu/Ha-NeRF/blob/7d722de779a6e8e526c38b0b384cc3f9a85e2ea5/losses.py#L55C68-L55C85