Can't find piece of code that normalizes predicted sdf to lie between -1 and 1

[Chonghyuk (Andrew) Song]

Hello,

I'm trying to follow up on the discussion in issue #12 about how the predicted SDF is defined (i.e. whether it is defined in the metric space or the normalized space [-1, 1]). You mentioned that "predicted SDF is always normalized between -1 (behind the surface) and +1 (in front of the surface)". However, I couldn't find anywhere in the code where the SDF output from the NeRF MLP is explicitly constrained to fall within the range [-1, 1] (for e.g. via a tanh activation) - there does not seem to be any activation applied when computing alpha_out in the function init_nerf_model from

neural-rgbd-surface-reconstruction/nerf_helpers.py

Lines 106 to 116 in 44ccc8a

 if use_viewdirs: 
     alpha_out = dense(1, act=None)(outputs) 
     bottleneck = dense(256, act=None)(outputs) 
     inputs_viewdirs = tf.concat([bottleneck, inputs_views], -1)  # concat viewdirs 
     outputs = inputs_viewdirs 
     for i in range(1): 
         outputs = dense(W//2)(outputs) 
     outputs = dense(3, act=None)(outputs) 
     outputs = tf.concat([outputs, alpha_out], -1) 
 else: 
     outputs = dense(output_ch, act=None)(outputs) 

.

Could you please direct me to where exactly in the codebase the predicted SDF is normalized?
Thank you in advance!

[Dejan]

There is no explicit normalization. The freespace and truncation losses push the predicted SDF towards the [-1, 1] range.