sbodenstein

5 ms·

1. 1.▲
   by sbodenstein 11y ago

   The Caffe models are only trained on the 1000 category subset of ImageNet used for the competition: http://image-net.org/challenges/LSVRC/2014/browse-synsets There are no sofa's in this list, the closest
2. 2.▲
   by sbodenstein 11y ago

   Standard convnets do not contain explicit rotational invariance (unless you include a layer such as this: arXiv:1506.02025v1). They can however learn rotational invariance if you feed them rotated images.
3. 3.▲
   by sbodenstein 11y ago

   We did not train ImageIdentify to be invariant under arbitrarily large rotations. This is fairly easy to do: show the network couches rotated at all angles.