Hungryof 于 2016.09.18 16:18 提问
I extracted features from last two conv layers from model VGG-19, separately named X and Y. X and Y share the same size 19*19*512. Then I apply dimension reduction to X and Y with PCA. And I get the size 361*20. 20 is the dimension.
Then I compute X's inner product, Y's inner product and the cross-product of X and Y, separately written XX,XY,YY. By now , XX ,YY or XY all share the size of 400*1. When I put XX or XY into linSVM as input, it shows badly efficiency and the accuracy of trainset is quite low(about 50%). However, YY as the input shows 100% accuracy of trainset and shows surprising speed. So it really confuses me. Is it possible?