Object Detection and Localisation in Images

by · Sep 25, 2014 · 2,271 views ·

Pražský informatický seminář

In the past 20 years, computer vision has been transformed from a purely theoretical subject producing algorithms failing on all but a few carefully chosen images into a large discipline where the path from a paper to a successful start-up company is surprisingly short. Multinationals like Google, Microsoft, Facebook, Adobe or Amazon employ dozens if not hundreds of computer vision experts. Detection and localisation of objects is one of the areas where significant progress has been achieved recently. It is now possible to answer, in near-real time, not only questions like "How many faces are there in the image?" but, for a broad class of objects, even "What is depicted in the picture?". Current research indicates there is no all-encompassing object detection algorithm, and a taxonomy of problems and approaches that has emerged will be introduced. I will present algorithms that have significantly pushed the state of the art and demonstrate the growing use of machine learning techniques on computer vision, focusing on the Viola-Jones detection method with sequential decision making. The talk will be concluded with remarks on results of the Bag of Words and Deep Neural Net methods.