search

Totoro97 / f2-nerf

Fast neural radiance field training with free camera trajectories
https://totoro97.github.io/projects/f2-nerf/
Apache License 2.0
933 stars 69 forks source link

about normal loss #71

Open xiepeter opened 1 year ago

xiepeter commented 1 year ago

I tried to add depth loss and normal loss like monosdf. It is quite smooth to add depth loss and it works for psnr to increase . but when it comes to normal loss: "Hash3Danchored" module can not handle gradient to wrapped points, because its gradient is for featpool. I tried to concat tcnn multi-hash embedding and sin/cos embedding to "Hash3Danchored" hash embedding, and use torch::autogrid::grad to calculate normal, but since the absence of "Hash3Danchored" loss, the total gradient is not complete. by some reason, the psnr dropped by 0.5 when incomplete gradient added. Would u give me some suggestion? thank u very much!

meriemjabri commented 1 year ago

could you please share how you added the depth loss and if you used your own point cloud how did you add it? I will appreciate any help!