Organisers: Antoni Buades (Universitat de les Illes Balears) and Gabriele Facciolo (ENS-Cachan)
Stereo vision attempts to estimate depth from two images. For thirty years many techniques have been put forward to solve this problem, many of which have been and remain successful in certain cases. However, general purpose high precision reconstruction is still elusive. New approaches such as active sensing, multiple viewing, and video processing are now being looked at. Thus the time is ripe to organize the main ideas behind two-image stereo vision. In this mini-symposium we bring together leading researchers to shed light on the current thinking in this area and to discuss new developments and challenges.
May 12, 2014. 5:15 PM - 7:15 PM (WLB208).
Hong Kong Baptist University, Hong Kong.
Kuk-Jin Yoon
School of Information and Communications,
Gwangju Institute of Science and Technology (GIST), Republic of Korea
Abstract: Even though a huge amount of researches have been done on two-frame stereo for last decades, we still have some questions about the two-frame stereo not clearly answered yet. Especially, when dealing with real-world images contaminated by several factors, the performance of the stereo methods is not acceptable. In this talk, I will discuss the fundamental limitations and difficulties of two-frame stereo in different aspects and discuss the possible research directions to overcome them.
Neus Sabater, Valter Drazic, Guillaume Boisson, and Paul Kerbiriou
Technicolor Research and Innovation, France
Abstract: In order to generate depth maps for professional applications such as Cinema we have designed a rig consisting of a professional camera with two HD cameras and a Kinect. A novel fusion algorithm combining view matching and depth sensing is proposed, generating fast, high-quality depth maps (reliable in textured and uniform areas). Our implementation generates quarter pixel accurate depth maps for HD-720p at 15fps. The output is suitable for virtual view synthesis and 3D reconstruction.
Jean-Marc Delvit
Centre national d’études spatiales (CNES), France
Abstract: The knowledge of ground elevation is essential in most remote sensing applications especially for very high resolution images. This ground elevation information can be retrieved from a pair of stereoscopic images, by correlation methods. The improving resolution of Earth observation systems like PleiadesHR (70 cm GSD) and their increasing stereoscopic capabilities open up new horizons for automatic Digital Elevation Model generation and allow to consider buildings reconstruction with an accuracy better than one meter RMS.
Gabriele Facciolo1 and Antoni Buades2
(1) CMLA, Ecole normale supérieure de Cachan, France
(2) TAMI, Universitat de les Illes Balears, Spain
Abstract: Motivated by photogrammetric applications we consider two issues of non-dense block-matching for stereovision. Correctly computing disparities at places where the pronto-parallel hypothesis is invalid (like discontinuities and slanted surfaces) and detecting invalid or ambiguous matches. We explore the use of oriented windows in order to deal with the first issue. While, to detect the mismatches we propose to use classic parameter-less techniques. Experiments show that incorporating these simple ingredients into a coarse-to-fine algorithm yield competitive results.