Build a convolutional neural net for image similarity

We will provide the following data: 1) Celebrities dataset: 800 images of celebrities 2) Dresses dataset: 10,000 images of dresses 3) Image similarity ground truth data: for each image in the celebrities dataset, a corresponding list of 10 image IDs referring to images from the dresses dataset (the most similar dresses that based on the associated celebrity image) We would like an image matching algorithm using neural nets (code and parameters) that for each input image in the celebrities dataset outputs a list of 10 image IDs referring to the dresses dataset (output is of the same structure as the Image similarity ground truth data) Project completion requirement: The accuracy of the algorithm should be of at least 80% measured as follows: - run the algorithm on a batch of 200 new celebrities images that we do not share (test dataset) - for each image, count how many of the 10 recommendations outputted by the algorithm are in the image similarity ground truth data for that input image - the average of these counts should be at least 8