Questions about camera parameters and scene parameter in dataset about hypernerf HOT 2 CLOSED

Hi Dogyoon. Thanks for your efforts!

What does the scale, scene_to_metric and bbox in scene.json file??

You can safely ignore scene_to_metric and bbox, they're not used. scale is the scale factor that converts the COLMAP coordinate to something smaller that can fit in a reasonable range. It's used here

Undistortion in LLFF vs here.

The difference is that some other NeRF papers pre-undistort the image and then use a pinhole camera, whereas we use the tangential and radial distortion parameters to generate the rays. The disadvantage of pre-undistorting is that you have to discard the border pixels since it can leave some invalid pixels. We also get to use the image pixels directly which can be nicer.

NDC

We don't use NDC, since some of our captures aren't fully front-facing and we don't really care about reconstructing far away backgrounds.

from hypernerf.

@keunhong hi, thanks for your awesome work! I have encountered a problem when I want to preprocess data.
I notice that the 'scale' in 'scene.json' of all your release data are the 0.04. But, I have preprocessed the original image sequences from 'rgb/1x' folder of each example, but the 'scale' of scene has a huge difference between the released data. And if I use my preprocessed data to train the nerf, it will get a bad result. I want to know the reason.
Thanks

from hypernerf.