Computer Vision News - September 2016

30 Computer Vision News Research Research Learned Invariant Feature Transform (LIFT) Every month, Computer Vision News reviews a research from our field. This month we have chosen to review the Learned Invariant Feature Transform (LIFT) , a research paper introducing a novel Deep Network architecture that implements the full feature point handling pipeline: detection, orientation estimation and feature description. The paper, which will be presented at ECCV 2016 in Amsterdam , shows how to learn to do all three in a unified manner while preserving end-to-end differentiability. This Deep pipeline model outperforms state-of-the-art methods on a number of benchmark datasets, without the need of retraining. We are indebted to the authors (Kwang Moo Yi, Eduard Trulls, Vincent Lepetit and Pascal Fua) for allowing us to use their images to illustrate this review. The full paper is here and the source code will be available here . Background: Finding and matching image features is a key step in many computer vision applications. The amount of literature relating to local features extraction is enormous, but it always revolves around three main steps: (a) finding feature points, (b) computing their orientation, and (c) describing them. Prior to the deep learning revolution, the best techniques relied on carefully hand-crafted image features. However, Deep Learning based techniques have started to outperform these traditional methods. Motivation: So far, all deep learning methods address only a single step in the feature extraction process (a to c above) and there is no existing method tying all three components together. Challenge: To train a Deep Network architecture that implements the full feature point handling pipeline: detection, orientation estimation and feature description. Novelty: A new framework for image features extraction called LIFT , which stands for Learned Invariant Feature Transform , the architecture that implements the full feature point handling pipeline (detection, orientation and description) is based “ Deep Network architecture that implements the full feature point handling pipeline, that is: detection, orientation estimation and feature description ”