The first layer only contains basic features, which are shared among different sets of images[1]. It seems that the ImageNet photos themselves are indeed very p

The first layer only contains basic features, which are shared among different sets of images[1]. It seems that the ImageNet photos themselves are indeed very protected[2], so I should probably use another dataset as input. Plugging in my iPhone camera roll should result in very similar filters. It will be very interesting to see how much they differ. Update: The author of Caffe just got back to me with word that the filters are in public domain, therefore safe to use as design elements. [1]: http://i.stack.imgur.com/Hl2H6.png http://i.stack.imgur.com/Hl2H6.png [2]: http://www.princeton.edu/main/administration/legal_compliance/copyright/ http://www.princeton.edu/main/administration/legal_complianc...