Qusetion about the Caluculation of the Mahalanobis Distance

[MayuOshima]

Qusetion about the Caluculation of the Mahalanobis Distance

[gengshan-y]

Hi, the code is cleaned up a bit. Hopefully it's more clear.

banmo/nnutils/geom_utils.py

Lines 232 to 234 in 6ce2b0c

	# multiply 1000 to make the weights more concentrated
	# log_scale controls global temporature scaling
	mdis = (-1000 * log_scale.exp() * mdis.sum(3)) # bs,N,B

[MayuOshima]

Thank you for your quick reply!
But I still have three questions, sorry.

1. How did you find the number 1000 to make the weight more concentrated? I mean, why you didn't use other numbers like 100, 2000, for example;
2. What is the global temporature scaling? I have Googled that but it seems there is not enough information about this term.
3. Given a mesh, how to get log_scale? If I guess right, it is an inherent attribute of a given mesh, right?

[gengshan-y]

Thanks for the follow up.

banmo/nnutils/geom_utils.py

Lines 232 to 235 in 2615647

	# log_scale (being optimized) controls temporature of the skinning weight softmax
	# multiply 1000 to make the weights more concentrated initially
	inv_temperature = 1000 * log_scale.exp()
	mdis = (-inv_temperature * mdis.sum(3)) # bs,N,B

To understand the temperature parameter better, perhaps this will help

1. The factor 1000 is empirically chosen by looking at the resulting skinning weights. 1000 produces skinning weights that has a clear separation between bones. This gives a good starting point for optimization.
   2-3. log_scale is a scalar parameter initialized as 0 and converges to ~1.5.

[MayuOshima]

OK! I totally understand that!
Thanks for your clearly reply!