10/25/2016: update the Demo code to clarify the hyper-parameter range
8/19/2016: correction for SynC.zip
Zero-shot learning codes & data
You can now download the codes (SynC.zip): https://www.dropbox.com/s/y7c03sr3vfdzry1/SynC.zip?dl=0
You can also download a baseline method ConSE (ConSE.zip): https://www.dropbox.com/s/oltpprntkjjspgn/ConSE.zip?dl=0
We provide the GoogLeNet features for three datasets (AwA, CUB, and SUN) in SynC.zip.
Note that on SUN, which has 717 classes in total, we split into ~646 seen classes and ~71 unseen classes, following Lampert et al., PAMI 2014. This is a much harder task than some other work that tests on only 10 unseen classes. The results of our method on such a simpler task are in the Suppl., where we have accuracy around 90%.
Feel free to contact me and let me know if you have any question about the codes. (Email: email@example.com)
Here we provide extra details for our CVPR 2016 paper:
Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha, "Synthesized classifiers for zero-shot learning," CVPR, 2016.
You can find links to our paper and the supplementary material via
Class splitting and semantic information for ImageNet
We provide the details of classes (seen, 2-hop, 3-hop, All) for the ImageNet zero-shot learning task as well as the word vectors in ImageNet_w2v.zip
You can also download it from https://www.dropbox.com/s/f9p3u6883xmt58k/ImageNet_w2v.zip?dl=0
Features for ImageNet
For extracting the GoogLeNet features, we follow:
The name of the network should be [CAFFE_ROOT] + 'models/bvlc_googlenet/bvlc_googlenet.caffemodel'.
Note that for the 1000 seen classes, we extract features from images in ILSVRC2012. For the 20842 unseen classes, we extract features from images in ImageNet2011release.