Monday 16:10–16:40 in Main Track

Spot the difference: train your image analytics model to recognize fine grained similarity

Katarina Milosevic, Ioana Gherman, Katarina Milosevic

Audience level:
Intermediate

Description

Imagine two images of the same car model, same color and with small scratches on bumpers. How would you make a machine look at the scratches and decide if they are the same? This is a story about the implementation of a new structure of neural network trained on “triplets” of images which recognizes fine grained images similarity based on Deep Ranking algorithm.

Abstract

Imagine two images of the same model of car, the same color and with small scratches on the bumpers. How would you make a machine look at the scratches and decide if they are the same? This talk will be a story about the implementation of a new structure of neural network trained on “triplets” of images. We will guide you through the implementation of the Deep Ranking algorithm for recognizing fine grained similarity of images, based on a real use case of car damages, with the purpose of fraud fighting. Presentation will follow the outline: • Business problem • Data description (images and metadata) • Triplets creation • Deep Ranking model • Results. The goal is to understand how is it possible to train a deep ranking model such that it focuses on damage differences. The training set contains positive pairs represented by pictures of a particular damage on the same car, and their corresponding negatives represented by very similar damages on the same car model. This network is able to learn similarity better and faster than methods based on vector comparison. The talk targets all Data Scientists interested in the Image Analytics topic, who are familiar up to a certain degree with the techniques already used in this field.

Subscribe to Receive PyData Updates

Subscribe