eiknoal_loss is an important term for SDF fields (e.g. NeuS / VolSDF).
For the early version (like v3.5), nerfacc operated the process in a fine-grined manner (ray marching sampling/field forward/volume rendering manually).
In this version (v5.2), nerfacc redesigned the APIs and packed these process in a pipeline consisted of estimator and rendering. The estimator design is really fantastic, but the rendering means that I could only consider my field (or a forward function) as an sigma&rgb generator.
In this way, once I want to get the points to calculate eikonal_loss, I need to record the interval, recover the sample points and feed them into the network to forward again.
Can the future version provide a more convient and flexible APIs? Here is just a friendly tip. :)
And the following is the example that I modify render_image_with_occgrid in utils.py from your examples to support the eikonal_loss calculation:
Thanks for your amazing project!
eiknoal_loss
is an important term for SDF fields (e.g. NeuS / VolSDF). For the early version (like v3.5),nerfacc
operated the process in a fine-grined manner (ray marching sampling/field forward/volume rendering manually). In this version (v5.2),nerfacc
redesigned the APIs and packed these process in a pipeline consisted ofestimator
andrendering
. Theestimator
design is really fantastic, but therendering
means that I could only consider my field (or a forward function) as an sigma&rgb generator. In this way, once I want to get the points to calculateeikonal_loss
, I need to record the interval, recover the sample points and feed them into the network to forward again.Can the future version provide a more convient and flexible APIs? Here is just a friendly tip. :) And the following is the example that I modify
render_image_with_occgrid
inutils.py
from yourexamples
to support theeikonal_loss
calculation: