Stereo reconstruction with Top-Down Segmented Regression

TDSR (Top-Down Segmented Regression) provides a stereo reconstruction method able to estimate the topography from SEM images. Standard stereo methods fail to evaluate adequate 3D reconstructions because of the homogeneous surface of these samples. The main approach of TDSR is to combine existing stereo methods with hierarchical segmentation of the images. Indeed, Mathematical Morphology provides efficient tools that divide an image into regions and subregions.
Instead of estimating an height map, TDSR refines and completes an existing height map obtained in a preliminary step. The reference image is first divided into main regions. For each region, the values of the initial height map inside the region are retrieved and modeled by a plane. If the model is satisfying, TDSR stop here. If it is not, the region is divided into subregions are modeled and followed the same process. This approach produces 3D reconstructions that are less noisy and more accurate than state of the art stereo methods for low texture SEM images.

This work was done during the phD thesis of Sébastien Drouyer, under the supervision of Serge Beucher* and Michel Bilodeau*, Maxime Moreaud+* and Loic Sorbier+. Sébastien Drouyer now works at CMLA°.
* MINES ParisTech, PSL-Research University, CMM, 35 rue Saint Honoré, 77305, Fontainebleau, France
+ IFP Energies nouvelles, Rond-point de l’échangeur de Solaize, BP 3, 69360 Solaize, France
° Centre de Mathématiques et Leurs Applications, École normale supérieure Paris-Saclay, 61, Avenue du Président Wilson, 94235 Cachan Cedex, France

If you use this plugin, please cite :
S. Drouyer, S. Beucher, M. Bilodeau, M. Moreaud, L. Sorbier. Sparse Stereo Disparity Map Densification Using Hierarchical Image Segmentation. International Symposium on Mathematical Morphology and Its Applications to Signal and Image Processing (ISMM) p. 172-184, Fontainebleau, France, 2017.