Search can be performed in various ways and using various architectures. In the following repository embeddings are created by using an non-cnn base embedding model which is tuned using a Siamese Neural Net with Triplet Loss inspired from F. Schroff, D. Kalenichenko and J. Philbin, "FaceNet: A unified embedding for face recognition and clustering," 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Boston, MA, 2015, pp. 815-823. paper from Google , on 10K MNIST digits images from train-set.The trained model is used to create embeddings for entire train-set and then Indexed using Approx Nearest Neighbours being faster than K-NN. More like a prototype.
PCA used for visualizing the embedding space of 10K images of different classes.
