search

shamangary / awesome-local-global-descriptor

My personal note about local and global descriptor
643 stars 94 forks source link

Questions about Matching loss for affinity matrix #13

Closed ghoshaw closed 4 years ago

ghoshaw commented 4 years ago

Hi, I have some questions about the Matching loss for affinity matrix. The affinity matrix M in your paper image, With my understanding，its shape is (HW)(HW), because each point in Fa must interact with every point in Fp to find the max corresponding. So, Fa is (HW)C, and Fp is C(HW). The loss Sa,p is calc by finding all the max corresponding for the two featrure map. But image confused me, in D2-net, the soft detection score M(det) seems to be HW*C, so which part is wrong?

shamangary commented 4 years ago

Hello @ghoshaw, thank you for your interest for our work. The detection score from D2-Net can be reduce to channel=1 by selecting the maximum value or taking the norm of the descriptor feature map, after that we also computer the affinity of the detection score therefore it's also (HW x HW).