How is Hamming Distance Used as a Loss for Identity?

[Atabuzzaman1]

Hi,

Thank you for the great work. Could you please provide more details on how you use the Hamming Distance (HD) as a loss to preserve identity during training?

[Siamul]

Hi, we don't use Hamming distance. If you binarize the filter outputs, it is no longer differentiable. What we do is run the filter outputs through sigmoid and then compare them using MSE or L1 or SmoothL1 loss. In a sense, you could think of this as an approximation of Hamming distance.

[Atabuzzaman1]

Got it. Thank you so much.