nn.init.orthogonal_ get better results than other initialization.
Another difference from netvlad is that only self.conv.weight is used for assignment and residual calculation, while the self.centroid is not. This comply with original VLAD that same visual words are used for assignment and residual calculation
sherlockers
commented
2 years ago
nn.init.orthogonal_ get better results than other initialization. Another difference from netvlad is that only self.conv.weight is used for assignment and residual calculation, while the self.centroid is not. This comply with original VLAD that same visual words are used for assignment and residual calculation