search

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

About DPT depth scale and shift initialization #4

Open K-nowing opened 1 year ago

K-nowing commented 1 year ago

Hi, thank you for code release

In the default config file, you used checkpoint "dpt_hybrid-midas", but why did you use the scale and shift value of checkpoint "dpt_hybrid-nyu" for preprocessing? I think the initial values of DPT depth affects the performance.

bianwenjing commented 1 year ago

Hi, thanks for your interest and thanks for identifying this problem. For the DPT model, we simply took the first model provided in their code and use a set of parameters they provided that can recover depth. It could probably be better if you use the corresponding fine-tuned checkpoint instead. You can also try other depth estimations as long as the depths are in reasonable scales within the sampling range.