This is a visual search demo. Click on a car, and the other cars will re-order by similarity. This is done by measuring the pair-wise euclidian distance between the feature vectors for the images. We generate these feature vectors using a pre-trained TensorFlow implementation of Google's Inception-v3 model. Although the Inception-v3 model is originally trained for a general 1000-label classification task, I postulate the second-to-last layer containing a 2048 dimension feature vector feeding into the softmax classifier can be used as a generalizable representation of an image (a.k.a. transfer learning). To learn more about the underlying convolutional neural network architecture for Inception-v3, check out this paper by Szegedy et al. (2015).
Find the code for this demo on Github.