Filmed at https://2017.dotai.io on April 25th in Paris. More talks on https://dotconferences.com/talks Image classification reached a new high thanks to advances in Deep Learning, by training a system at tagging images thanks to millions of examples. But understanding what's in images in the wild remains a challenging task: it requires identifying the different objects, where they are located in the image and how to affect precise tags to these objects. Charles shows how we can strongly reduce the number of training images and maintain good performance over wide ranges of inputs. He showcases the pipeline, tools and methods developed at Heuritech for tackling these challenges and build a large scale computer vision pipeline.