This visualization shows 1000 samples from the Women's clothing category of the Amazon clothing catalog. Product images are fed thought a Convolutional Neural Net where the 4096 dimensional visual features are taken at the output of the second fully-connected layer (i.e., FC7). The CNN is trained is a caffe reference model trained on imagenet. Tsne is then used to project the image features down to 2D. PCA preprocessing is used prior to the tsne routine to reduce to 10D to help optimize the tsne runtime.

zoom and scroll to explore relationships