Carmelo VELARDO - PhD student Multimedia Communications
Date: - Location: Eurecom
Face Recognition with DAISY Descriptors In this work we propose a new face recognition approach based on DAISY, a dense computed SIFT-like descriptor. Our algorithm is designed to be fast for dense computation, and useful for re-identification as it is able to distinguish pairs of images as belonging to the same subject or not. The descriptors are computed densely and matched with a new strategy that represents an efficient trade off between accuracy and computational load; afterwards a Support Vector Machine is used to classify the output of the matching to recognize if the pairs of images belongs to the same person. An analysis of performance will be conducted on two different databases in order to compare our results with the already existing ones. We show that better performance than SIFT techniques can be achieved using our algorithm. ******* Weight estimation from visual body appearance Weight is a biometric trait which has been already studied in both the forensic and medical domains. In many practical situations, such as videosurveillance, weight can provide useful information for re?identification purposes but needs to be estimated from visual appearance (images or video). In this work we study the feasibility of weight estimation from anthropometric data directly accessible from the available image material. A model is retrieved via multiple regression analysis on a set of anthropometric features. A large medical database is exploited for the model training, while its validation is performed both on ideal and realistic conditions. The performance analysis of our approach shows that under noisy data conditions the system provides accurate estimations, putting the basis for a future work towards an automatic weight estimation.
