Description
Hello, Thanks for your work.
I see that on the Test Data you do L2 Norm to the features is that correct? Wouldn't that throw off the threshold value since it was learned on an unnormalized d^2 euclidean space instead of normalized cosine space?
Also, I was trying to implement the BallClustering loss myself. Is it the same to assume its similar to CenterLoss but instead of the loss being the distances to the Centroid we use the 2 d^2 euclidean constrains to calculate loss? As parameters for the loss I have the centroids (num_classes, num_features) and b_hat (scalar value). It doesn't seem to produce better results than Softmax + Triplet on my datasets can you help me out figuring out what's missing or share your BallClustering Loss to compare with my implementation?
In the paper, you mention the Loss is simply L = alpha * Lsim + Ldis but no mention of Softmax. Is it safe to assume it's both Softmax + BallClustering on just 1 loss?
Thanks.