Skip to Main content Skip to Navigation
Conference papers

Apprentissage de la Cohérence Photométrique pour la Reconstruction de Formes Multi-Vues

Vincent Leroy 1 Jean-Sébastien Franco 1 Edmond Boyer 1
1 MORPHEO - Capture and Analysis of Shapes in Motion
Inria Grenoble - Rhône-Alpes, LJK - Laboratoire Jean Kuntzmann, INPG - Institut National Polytechnique de Grenoble
Abstract : With the rise of augmented and virtual reality, estimating accurate shapes from multi-view RGB images is becoming an important task in computer vision. The dominant strategy employed for that purpose in the recent years relies on depth maps estimation followed by depth fusion, as depth maps prove to be efficient in recovering local surface details. Motivated by recent success of convolutional neural networks, we take this strategy a step further and present a novel solution for depth map estimation which consists in sweeping a volume along projected rays from a camera, and inferring surface presence probability at a point, seen by an arbitrary number of cameras. A strong motivation behind this work is to study the ability of learning based features to outperform traditional 2D features when estimating depth from multi-view cues. Especially with real life dynamic scenes, containing multiple moving subjects with complex surface details, scenarios where previous image based MVS methods fail to recover accurate details. Our results demonstrate this ability, showing that a CNN, trained on a standard static dataset, can help recovering surface details on dynamic scenes that are not visible to traditional 2D feature based methods. In addition, our evaluation also includes a comparison to existing reconstruction pipelines on the standard evaluation dataset we used to train our network with, showing that our solution performs on par or better than these approaches.
Complete list of metadatas

Cited literature [42 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-01857627
Contributor : Vincent Leroy <>
Submitted on : Friday, August 17, 2018 - 11:10:26 AM
Last modification on : Tuesday, April 14, 2020 - 5:14:02 PM
Document(s) archivé(s) le : Sunday, November 18, 2018 - 12:36:07 PM

File

rfiap2018.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01857627, version 1

Collections

Citation

Vincent Leroy, Jean-Sébastien Franco, Edmond Boyer. Apprentissage de la Cohérence Photométrique pour la Reconstruction de Formes Multi-Vues. RFIAP 2018 - Reconnaissance des Formes, Image, Apprentissage et Perception, Jun 2018, Marne la Vallée, France. ⟨hal-01857627⟩

Share

Metrics

Record views

221

Files downloads

187