search

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

how is the number of channels changed from 96 to 48 as input of conv2? #24

Closed xiaowei-hu closed 7 years ago

xiaowei-hu commented 7 years ago

how is the number of channels changed from 96 to 48 as input of conv2?

The number of the feature channels of conv1 is 96. However after batchnorm, relu and pool layers, it changes to 48. This confuses me a lot. Which layer changes the number of channels?

Jiangfeng-Xiong commented 7 years ago

As I know, it is the same as Alexnet. If you know caffe, it means that we set group to 2 in Convolutional Layer.That is to say

the first half of the filters are only connected to the first half of the input channels, and the second half only connected to the second half.

So,it's half of the original input channel num 96, (namely,48). In conv2,conv4,conv5, It is the same. you can refer to this issue of caffe

xiaowei-hu commented 7 years ago

thx, i got it.