Data-driven detection and self-learning camera
Place: Large Lecture Room
Affiliation: Xerox Research Centre, Grenoble, France.
The first part of the talk will cover "Data-driven detection": an approach to detect prominent objects in images WITHOUT using sliding windows, just computing a single feature vector per image. We formulate object detection as a retrieval problem: given a query image, retrieve images with the same layout, and output a simple combination of these layouts. The key element of our approach is a "learning to rank" framework that is designed to explicitly optimize detection accuracy, and a probability map representation of images. This is joint work with Diane Larlus.
The second part will discuss a work dubbed "Self Learning Camera". The goal is to detect objects in videos with ZERO labelled images of the target video. We employ an unsupervised, continuous domain adaptation approach that starts with a pre-trained detector and automatically derives a better detector for the target video. The algorithm exploits an ensemble of trackers, with capabilities to mine hard negatives, and a mean-regularized multi-task learning framework to derive a class-level object detector from the trackers. This is joint work with Adrien Gaidon, Eleonora Vig and Gloria Zen.