Difference between EMD loss and smoothL1?

[ming71]

Your implementation of EMD loss here seems to be the same as smoothL1 loss between anchor and its pred boxes.What's the difference them?Maybe I don't figure it out clearly..

[xg-chu]

The EMD Loss calculates the minimum loss between the two sets. The single loss between each element in the two sets is measured by SmoothL1 and softmax Cross Entropy.

[ming71]

For each element in anchor set , two boxes will be regressed, then you apply smoothl1 and cross entropy to it ....It's almost the same as traditional loss do.In this way , I can also regard traditional single loss as EMD Los between anchor set and pred_boxes set while num of pred_boxes is equal to that in anchor set , right?

[ming71]

In details , given anchor set A in shape of (1, 1, 4) (bs=1, num_anchor=1, 4=xywh),and pred boxes set B in shape of (1, 2, 4), when we calc EMD loss for regression, we can obtaion res = smoothl1(A[0,0], B[0,0]) + smoothl1(A[0,0], B[0,1]) .
for single loss, pred boxes set B will be in shape of (1, 1, 4), and result will be res = smoothl1(A[0,0], B[0,0]) .

sorry to disturb you, is there something wrong with my understanding?Thank you.

[xg-chu]

The anchor is regressed by two parallel prediction head. The anchor A is predicted as P_head0 and P_head1.
The loss is between P_heads and ground truths.
The EMDLoss calculate the min ((L(p_head0, B0) + L(p_head1, B1)),(L(p_head0, B1) + L(p_head1, B0))). Ensure that the network can always be optimized in the optimal direction.

[ming71]

thanks for your reply, I make it to understand now