b



Member of Titane
Geometric Modeling of 3D Environment


Mohammad Rouhani, PhD

Postdoctoral Researcher

Titane team, INRIA Sophia-Antipolis
06420 Nice, France
www.cvpr3D.com
mohammad.rouhani [at] inria dot fr
Tel: 
+33-497155309

b
Home Research Publications Projects Students
b

  
   Please refer to the latest version of my CV.

I have been working on different topics of 3D Computer Vision and Graphics, including surface reconstruction and representation, surface and volume registration, shape matching as well as modern Machine Learning for object detection and semantic segmentation. During my PhD I had the chance of working with Angel D. Sappa at Computer Vision Center. My research skills have been further elaborated during my Postdoc opportunities at Imperial College in London and INRIA Rhone-Alpes in Grenoble. At the moment I am with INRIA Nice working with Pierre Alliez and Florent Lafarge on semantic segmentation of 3D textured meshes.


The list of topics I have been working on are listed bellow:

b


Given a textured mesh (in *.OBJ) it must be semantically segmented into pieces with the right labels. For this purpose a modern machine learning techniques is employed on a set of training examples described by photometric and geometric features...





Volumetric Shape Matching:

Traditional matching methods are based on the surface manifold. By exploiting the volumetric information, we model the matching as a mass transportation problem while the cell's volumes are preserved. The result is followed by a volumetric deformation that is more natural for solid objects. 









Multi-Class Object Detection:

Randomized Decision Trees are among fast and popular tools in machine learning and computer vision. We use this tool for multi-class object detection as well as 3D pose estimation. The employed random forest is based on dominant orientation templates, instead.



 

Deformation Modeling:

This is a common field of study between computer vision and graphics. Deformation modeling concerns transforming the given image or mesh so that its basic properties are preserved during the transformation. Based on the criteria used as a shape prior the deformation might be intrinsic or extrinsic and the optimization framework may vary depending on the model. I have a long experience working with FFD models and as-rigid-as-possible approaches where a simple least square can be obtained to preserve the shape prior.


b



Non-Rigid Shape Registration:

The registration distance original proposed for rigid case is extended for non-rigid deformation. The distance is robust to noise and missing data, hence it shows a good behavior for non-rigid registration as well. The deformation space could be selected as TPS or iFFD or even the Laplacian Deformation.

b


2D and 3D Point Set Alignment:

In this work, we convert the point-to-point alignment to a point-to model alignment problem. We used implicit representation to describe the model set. Then we would be able to exploit many metrics provided by this representation. As a result the objective function will be a smooth function which can be used in many gradient based optimization algorithms.



b