Feature dimension of final classifier

[insomnia250]

@wangzheallen Thanks for sharing your interesting work! I have a little question about the feature dimension of the final SVM classifier. According to your paper and code, the dimension of vsad code ( i.e. the features for SVM classifier) is 2*len(f)*len(p) , where f denotes descriptors from your scene-PatchNet's feature layer(with dimension of 100 reduced from1024 ) and p denotes codewords from object-PatchNet's softmax layer(i.e. probabilities with dimension of 256 reduced from 1000 ). Even if dimension reduction is employed, the feature dimensions for SVM are still very high(say, 2×100×256=51200 according to your paper). Will this cause some problems to the final classification? What do you think? (Let me know if i was wrong somewhere! ) Thanks in advance!

[wangzheallen]

Hi insomnia250,

thanks for your interest in our work! Based on my experiment I have not found anything wrong with the performance with such high dimension. But I never tried to use PCA directly on the '51200' dimensional feature.

And I list the paper with high dimensional feature and get very good performance http://www.sciencedirect.com/science/article/pii/S1077314216300091 http://xjpeng.weebly.com/uploads/5/5/4/4/55444193/th15_inria.pdf http://xjpeng.weebly.com/uploads/5/5/4/4/55444193/pwqp_eccv14_shvlad.pdf

Zhe

[insomnia250]

@wangzheallen OK, I guess I should first try it following your paper any way :). Thanks for your reply!