Skip to Main content Skip to Navigation
Theses

Méthodes d'apprentissage pour l'estimation de la pose de la tête dans des images monoculaires

Abstract : This doctoral research is part of PILE, a medical project which aims at analyzing baby's gazes, gestures and vocalizations. In this context, we have designed and developed methods for determining the head pose which constitutes the cornerstone of a system for estimating the gaze direction. From a methodological point of view, we have proposed BISAR (Boosted Input Selection Algorithm for Regression), a feature selection method which is well adapted to regression problems. It consists in iteratively selecting inputs of an incremental neural network. Each input corresponds to a feature selected by our Fuzzy Functional Criterion. The latter measures the functional relation between a feature and the values to predict. The features complementarity is provided by a boosting process that changes weight distribution on the training examples. This algorithm has been experimentally validated in two head pose estimation methods. The first approach directly learns the relationship between the appearance of a face and its corresponding pose. The second approach aligns a face model in an image and then calculates the geometric orientation of this model. The alignment process is based on a cost function that evaluates the quality of the fitness. This function is learned by BISAR from examples of aligned and misaligned models. Evaluations of these methods have given state of the art results on different test sets with large variations in pose, identity, illumination and shooting conditions.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00560836
Contributor : Kevin Bailly <>
Submitted on : Sunday, January 30, 2011 - 8:32:15 PM
Last modification on : Thursday, January 23, 2020 - 12:04:08 AM
Document(s) archivé(s) le : Thursday, March 30, 2017 - 6:05:30 AM

Files

Identifiers

  • HAL Id : tel-00560836, version 1

Citation

Kévin Bailly. Méthodes d'apprentissage pour l'estimation de la pose de la tête dans des images monoculaires. Interface homme-machine [cs.HC]. Université Pierre et Marie Curie - Paris VI, 2010. Français. ⟨tel-00560836⟩

Share

Metrics

Record views

632

Files downloads

4466