[ Training Loss: Jensen Shannon Divergence ]

[IemProg]

Hi,

Great work, thanks for sharing the code !

I would like to apply your implementation of JSD, in another pose estimation framework. Could you please, share what's the shape of your spherical gaussians (_makegauss() function). Are you generating a gaussian distribution for each joint ?

In my case, heatmaps are structured such as: XY = [ N, Nbr_Joints, Height, Width ]. I'm expecting an output of _makegauss() to be [Nbr_joints, Height, Width] but i'm getting [N, Height, Width] ?

Thanks !

[anibali]

You must call make_gauss with coordinates representing the means (centres) of the Gaussians to draw. Consider the signature:

make_gauss(means, size, sigma, normalize=True)

Now, for your case,

• means is expected to be of shape [N, Nbr_Joints, 2], where the last axis stores coordinates in (x, y) order.
• size should be set to (Height, Width), which will be the spatial size of the generated Gaussian images.
• sigma should be set to the desired standard deviation (spread) of the Gaussian, in pixels.

The tensor returned by the function will have shape [N, Nbr_Joints, Height, Width].