Question about a value in MobileNeRF about jax3d

[flyingshan]

I think the world_tx in compute_undc_intersection denotes the T.

[G-1nOnly]

@flyingshan I think it somehow mismatches the "T" for the original value which lies in [near, far] but the value of world_tx could be very large sometimes, is it because of the sampling way is different?

[flyingshan]

I think you are right. I check the code again. The world_tx just indicates how far the intersection points are away from the origin. I think you can calculate the real T value from the ray_origin, ray_directions, and the world_positions through points = origin + direction * T.

[G-1nOnly]

Yeah, I've tried this way, but the "T" caculated seems not correspond to all the points by this way, I think it's because the uniqueness of the sampleing method since normally we get T, origin and direction to calculate points but MobileNeRF gets the points in its own way and when calculating the weights, MobileNeRF directly optimizes "alpha" instead of sigmas normally in NeRF.