Seems that the abs activation is redundant?

[yzslab]

Hi, I found that the opacities will not be optimized after replacing the activation function with the abs here, due to the absence of the replace_tensor_to_optimizer operation.
Since the range of the Sigmoid is $$(0, 1)$$ and the opacities are fixed, they will never fall below 0. Therefore, it seems that the abs activation is redundant, as simply returning the same tensor suffices.
I tested removing the abs by the modification below and it produced the same metrics:

diff --git a/scene/gaussian_model.py b/scene/gaussian_model.py
index 8c65c4c..0e46ba8 100644
--- a/scene/gaussian_model.py
+++ b/scene/gaussian_model.py
@@ -46,14 +46,14 @@ class GaussianModel:
             self.inverse_opacity_activation = inverse_sigmoid
         else:
             print("Absolute rendering mode")
-            self.opacity_activation = torch.abs
+            self.opacity_activation = identity_gate
             self.inverse_opacity_activation = identity_gate
 
         self.rotation_activation = torch.nn.functional.normalize
 
     def modify_functions(self):
         old_opacities = self.get_opacity.clone()
-        self.opacity_activation = torch.abs
+        self.opacity_activation = identity_gate
         self.inverse_opacity_activation = identity_gate
         self._opacity = self.opacity_activation(old_opacities)

Is this expected?

[saswat0]

@yzslab You are correct. The abs activation is kept just as a failsafe and we don't expect any performance improvements with it

[yzslab]

@yzslab You are correct. The abs activation is kept just as a failsafe and we don't expect any performance improvements with it

Hi, does the modify_functions correspond to the Section 3.4 of your paper?

Therefore, we convert the regular, capped Gaussian primitives to high-opacity Gaussians after reaching the midpoint of our training (15K iterations). This involves replacing the opacity activation with abs and clamping blending weights to 1 from above during rendering. As shown by our ablation, this change positively impacts quality metrics, particularly PSNR.

Could you please confirm if this conclusion is still valid?

[saswat0]

@yzslab Apologies for the late response.

Thanks for pointing out this issue. We've patched the bug with d1fced9

[yzslab]

@yzslab Apologies for the late response.

Thanks for pointing out this issue. We've patched the bug with d1fced9

Thanks, but have you tested it? Seems that such a simple solution can not produce the expected results.