bertinetto / siamese-fc

Arbitrary object tracking at 50-100 FPS with Fully Convolutional Siamese networks.
http://www.robots.ox.ac.uk/~luca/siamese-fc.html
MIT License
619 stars 224 forks source link

element-wise label size is 15*15 in the code while 17*17 in the paper #26

Closed NicoleWang closed 7 years ago

NicoleWang commented 7 years ago

In siamese-fc paper, after cross-correlation layer, the output is a 1717 array which means that the element-wise label should be a 1717 array too. But in the released code, I found out that the element-wise label size is 15*15. It's quite confused.

jvlmdr commented 7 years ago

The score map (and element-wise labels) has size 17x17 for a search region of size 255 pixels. However, to enable data augmentation, we instead use a search region of size 255-2*8 = 239 pixels. This reduces the size of the score map to 15x15, since the score map has stride 8 with respect to the search region.

NicoleWang commented 7 years ago

Got it. Thanks a lot ~