Version française
RSS feeds
Home > News > News stream... > Image processing: an innovative stereo 3D reconstruction approach

News stream...

Image processing: an innovative stereo 3D reconstruction approach

September 2017

As part of a thesis project and for the purposes of its research in the field of catalysts, IFPEN has co-developed a new stereo reconstruction approach from images obtained using an electron microscope. The method developed has also proved effective beyond this initial context, opening up avenues of potential interest in other fields of application.

Whether it be to determine the relief of urban zones, supply three-dimensional information - to robots or driverless vehicles or to characterize the surface of small industrial objects, such as catalyst grains, obtaining the 3D topography of objects or places - is increasingly being seen as an interesting approach in numerous fields.

Consequently, so-called stereo reconstruction methods, which provide a 3D representation from similarities between two (or more) images of the same scene, are attracting increasing attention. These methods are based on matching similar points between images. The differences in the positioning of these points - referred to as disparity maps - can be used as a basis for recreating 3D information concerning surfaces.

By applying these stereo reconstruction methods to images obtained from a scanning electron microscope in order to obtain a 3D reconstruction of granular materials such as refining catalysts (zeolite catalysts, for example), researchers have developed a new approach. The new approach utilizes an initial disparity map calculated using a standard method, then corrects it and enhances it.

What is innovative about the approach is the fact that it consists of two successive steps:

  • the robust, multi-scale segmentation of the image into small zones - making it possible to assemble correspondence points and eliminate abnormal disparities - followed by the propagation of this information to occluded zones (zones only visible on one of the two images);
  • then, for each of the zones, the setting of parameters for a geometric model (3D plans, for example) enabling an additional reduction of abnormal disparities.

This approach was the subject of an oral paper [1] presented at the recent 2017 ISMM conference.

This new approach was also transposed to satellite images as part of the Multi-View Stereo 3D Mapping Challenge organized by the American IARPA (Intelligence Advanced Research Projects Activity) agency, which consisted of two distinct contests: Explorer and Master. It was ranked first in the Explorer contest and third in the Master contest, as a result of which the team was invited to present its work at an IARPA conference held in Washington in November 2016.

The method can also be used on photographs of everyday scenes and has been evaluated on the Middlebury database (the benchmark stereo images database). It proved comparable to the state of the art, oscillating in the official ranking between 1st and 3rd place, depending on the criteria selected.

This research was carried out as part of Sébastien Drouyer’s thesis, supervised by Serge Beucher and Michel Bilodeau (Mathematical Morphology Center / PSL-MINES ParisTech), along with two IFPEN researchers: Loic Sorbier (Physics and Analysis Division) and Maxime Moreaud (Mechatronics, Computer Science and Applied Mathematics Division). The image processing algorithms developed during this thesis will be taken forward to a fundamental research project conducted at IFPEN, and will soon be available in an internal image processing platform, INDIGO.

Scientific contact:


  1.  S. Drouyer, S. Beucher, M. Bilodeau, M. Moreaud, L. Sorbier. Sparse stereo disparity map densification using hierarchical image segmentation. 13th International Symposium on Mathematical Morphology, Fontainebleau ( France), 15-17 Mai 2017.
    >> hal-01484143

links list

  • Print page

Follow IFP Energies nouvelles (IFPEN)

on Twitter

on YouTube