I just wanted to build something cool using machine learning on a bunch of public images. Essentially building a database of indexed faces ( ( ʖ( ʖ)ʖ)). The coin sorting machine can measure the coins' size and weight, and will sort coins of similar size and weight into one pile. FaceNet is doing exactly the same thing, except it's measuring the high level abstract features of the images (position of mouth, eyes, nose), and not low level concrete features (skin color, hair color, etc) Because of the Triplet Loss, it can be assumed that faces with similar features will have a similar embedding.
Jun-19-2017, 19:15:13 GMT