A Region-Based Randomized Voting Scheme for Stereo Matching

Abstract

This paper presents a region-based stereo matching algo- rithm which uses a new method to select the nal disparity: a random process computes for each pixel di erent approximations of its disparity relying on a surface model with di erent image segmentations and each found disparity gets a vote. At last, the nal disparity is selected by estimating the mode of a density function built from these votes. We also advise how to choose the di erent parameters. Finally, an evaluation shows that the proposed method is e cient even at sub-pixel accuracy and is competitive with the state of the art.

Slides

A Region-Based Randomized Voting Scheme for Stereo Matching from Guillaume Gales

Download

Download PDF.

Complementarity of feature point detection

Abstract

The goal of this paper is to provide a study on complementarity of feature point detectors for stereo image pairs. Many studies have been proposed on these detectors but none deals with complementarity in details. We introduce an evaluation of eleven well-known detectors (Moravec, Harris, Kitchen-Rosenfeld, Beaudet, SUSAN, FAST, SIFT, Harris-Laplace, Hessian-Laplace, SURF, Kadir) based on new criteria used to characterize complementarity. The complementarity is evaluated with: Spatial distribution, Contribution measures, Repeatability gain, Distribution gain.

Download

Download PDF
Download Slides (PDF)

Feature point detection for pixel matching by propagation

Abstract

In a binocular stereovision context, we focus on matching techniques based on seed propagation. Our work is about the seed selection step. Among all the different solutions, some are based on feature points. The goal of this work is to propose a study on ten feature point detectors in order to find out which ones are the most suitable to select the initial seed set for the propagation. In order to do so, we adapt the auto-correlation measure proposed by Moravec by using a different correlation measure over squared windows and by looking over a bigger neighbourhood. Several rankings are established measuring repeatability and spatial distribution computed from a set of reference stereo image pairs. This work gives indications about the most adequate strategy to select reliable and well distributed seeds. Then, we take a look at the complementarity of the most efficient detectors according to our criteria. Then, we take a look at the influence of the most complementary detectors in the stereo matching context.

Download

Download PDF (in french)