Question about the inconsistence of depth

ActiveVisionLab / nope-nerf

(CVPR 2023) NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior

https://nope-nerf.active.vision/

MIT License

372 stars 30 forks source link

Question about the inconsistence of depth #21

Closed lbh666 closed 11 months ago

lbh666 commented 12 months ago

Hi, thanks for your great work! And I have a question about the detail of your implementation. As shown in the attached image, the depth estimated by nerf usually refer to the range between object and camera. But the DPT depth output is the perpendicular distance between object and camera plane (the definition of depth). Did you consider the difference between them? Please correct me if I were wrong. Looking forward to your early reply. machines1000183g002-622803161b723

bianwenjing commented 11 months ago

Hi, sorry for my late reply. The difference between depth (perpendicular distance) and distance(the actual range) is considered. During training, the depth value of DPT is changed to distance at https://github.com/ActiveVisionLab/nope-nerf/blob/47c861f6259fb0f3921be3a19385feebfa769325/model/rendering.py#L60C12-L60C12. The ray vector of NeRF is also normalised, so they should be consistent.

lbh666 commented 11 months ago

Got It, Thanks!