Closed ch3cook-fdu closed 1 year ago
We use huber_loss to represent boundary_reg loss. The weights are set empirically. Be free to change it by yourself. https://github.com/zlccccc/3DJCG/blob/053ae910384d280cd4e4a4e473bfc003ab684d12/lib/loss_helper/loss_detection.py#L141
In page 16468 of your paper. Is the weight of
boundary-reg
loss way too large?$$ L{detection} = 10L{vote-reg} + L{objn-cls} + L{sem-cls} + 200L_{boundary-reg} $$