Représentations à base de parties pour la vision 3D de haut niveau

Abstract : In this work we use Deformable Part Models (DPMs) to learn and detect object parts in 3 dimensions. Given a single RGB image of an object, the objective is to determine the location of the object’s parts. The resulting optimization problem is non-convex and challenging due to its large solution space.Our first contribution consists in extending DPMs into the third dimension through an efficient Branch-and-Bound algorithm. We devise a customized algorithm that is two orders of magnitude faster than a naive approach and guarantees global-optimality. We derive the model’s 3-dimensional geometry from one 3-dimensional structure, but train viewpoint-specific part appearance terms based on deep learning features. We demonstrate our approach on the task of 3D object pose estimation, determining the object pose within a fraction of a second.Our second contribution allows us to perform efficient inference with part-based models where the part connections form a graph with loops, thereby allowing for richer models. For this, we use the Alternating Direction Method of Multipliers (ADMM) to decouple the problem and solve iteratively a set of easier sub-problems. We compute 3-dimensional model parameters in a Convolutional Neural Network for 3D human pose estimation. Then we append the developed inference algorithm as final layer to this neural network. This yields state of the art performance in the 3D human pose estimation task.
Document type :
Theses
Complete list of metadatas

Cited literature [200 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01885958
Contributor : Abes Star <>
Submitted on : Tuesday, October 2, 2018 - 2:33:16 PM
Last modification on : Tuesday, April 16, 2019 - 9:46:23 AM
Long-term archiving on : Thursday, January 3, 2019 - 3:00:20 PM

File

75422_KINAUER_2018_archivage.p...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01885958, version 1

Citation

Stefan Kinauer. Représentations à base de parties pour la vision 3D de haut niveau. Mathématiques générales [math.GM]. Université Paris-Saclay, 2018. Français. ⟨NNT : 2018SACLC059⟩. ⟨tel-01885958⟩

Share

Metrics

Record views

174

Files downloads

164