Does Scarlett smile more than her French lover ? Constraints and regularization in visual metric learning
From 03/02/2015 at 10:30 to 12:00. Amphi Emilie du Châtelet, bibliothèque Marie Curie, INSA de Lyon
URL : https://liris.cnrs.fr/seminaire/seminaires-mensuels/seminaires-mensuels
Informations contact : G. Damiand. firstname.lastname@example.org. +33 (0)220.127.116.11.34.
Metric learning is useful in many Computer Vision applications, such as image classification, retrieval, face verification, or person re-identification. Supervised metric learning has been deeply investigated to compare feature vectors. I will focus in this talk on Mahalanobis-like distance metric learning that essentially infer a linear transformation of the data space. The key components of this learning problem are data constraints and regularization.
I first present classic metric learning approaches based on image pairs or triplets constraints. I discuss specific extensions focusing on image attributes (e.g. smiling) ranked by classes (e.g. a given celebrity) or inferred from richer relationships, e.g., a hierarchy of classes. In a second part, I discuss the regularization term for this type of learning scheme. In particular, I present classic methods to minimize the rank of the learned model in a Mahalanobis-like matrix learning. I then introduce interesting regularizer to explicitly control the matrix rank. Several results on image datasets and Webpage screenshot comparison will be discussed.