search

zxhuang1698 / interpretability-by-parts

Code repository for "Interpretable and Accurate Fine-grained Recognition via Region Grouping", CVPR 2020 (Oral)
https://www.biostat.wisc.edu/~yli/cvpr2020-interp/
129 stars 27 forks source link

The Implementation of Region Feature Extraction #3

Closed maktu6 closed 4 years ago

maktu6 commented 4 years ago

Dose the implementation of Region Feature Extraction in here equal to Equation(3) in the paper, Can you explain it more concretely? And what is the difference of just using qx = torch.bmm(assign, x) as the out, dose the model performance would degrade?

zxhuang1698 commented 4 years ago

Yes, I think so. Please see the updated comments of that file. For computing the residual coding, note that the grouping centers do not need to go through the \sum{q_ij d} / \sum{q_ij} step because \sum{q_ij d} = \sum{q_ij} * d.

For residual coding, we follow the commonly used nonlinear feature encoding scheme. You can definitely try to remove the residual term, it might hurt the accuracy to some extent. But I think you still need the normalization term to make the magnitude consistent.

maktu6 commented 4 years ago

Thanks for your reply!