Average Hausdorff loss implementation?

1

Hi there,

does anyone of you know an open source implementation of the Average Hausdorff distance? I would like to try using it for some 3D volumetric data but don’t quite know how to get started with writing a loss function in pytorch.

I am helpful for any kind of input.

Thanks a lot.
Christian

1 Like
2

Hi Christian,

The open source code of a CVPR 2019 paper (Locating Objects Without Bounding Boxes) may be helpful. Here is the url.

Not sure whether this is still helpful you right now.

3

Thanks for this reference. It seem that it’s not applied to 3D. Am I correct?

4

5 years late, but here is a more generic implementation

X1.shape == (B, N, D)
X2.shape == (M, D)

X2 = X2.unsqueeze(0).expand(B, M, D)   # (B, M, D)
dist = torch.cdist(X1, X2_exp) # B, N, M 

dh_x_y = dist.min(1)[0].mean(-1) # B, M --> B 
dh_y_x = dist.min(2)[0].mean(-1) # B, N --> B 
dist = dh_x_y + dh_y_x