Learning Shape Topology

The purpose of this work is to sequentially learn the correct topology of a scene. Based on incomplete observations, such as a temporal sequence of meshes reconstructed from visual hulls, we can progressively build a template mesh that will fit the entiere sequence. Our method rely on the assumption that object have a fixed topology and do not physicaly merge. Hence every new topological evidence observed is added to the template.

Want to know more?

Will be filled after CVPR 2012

Hide this panel.

Scene Flow

In this work we study how to incorporate, in an efficient way, various constraints when estimating dense motion information over 3D surfaces from temporal variations of the intensity function in several images. Our primary motivation is to provide robust motion cues that can be directly used by an application, e.g. interactive applications, or that can be fed into more advanced tasks such as surface tracking or segmentation. The approach is however not limited to a specific scenario and applies to any application that can benefit from low level motion information. We explore the combination of heterogeneous photometric and geometric cues in an unified framework. We fuse 2D and 3D, sparse and dense information to constrain the estimation of the instantaneous 3D scene flow on a surface. This project uses the output of any multi-camera system, such as Grimage, which gives for each time step a set of calibrated images. Our framework merges constraints coming from 2D normal flow, sparse 2D features matching in the images, sparse 3D features matching on the surfaces and a regularization term to provide an accurate and smooth motion field over the surface.

Want to know more?

Here are some videos showing results on real datasets

Pipeline of our method for depht maps


Other materials

BMVC Poster BMVC Abstract VMV Slides

Hide this panel.

Master thesis - Plane detection in images

This is the work I did during my research internship during my second year of Master in Toulouse. The purpose was to use line correspondances in image pairs to detect simultaneously every plans in a scene. Fisrt I adapted an existing technic based on generalized principal component analysis developped by Yi Ma used to find the number and parameters of subspaces defined by data points in a space of higher dimension. The results were nice but very sensitiv to noise. That's why we decided to devellop a new approach, we used a technic that fits well in the context of urban scenes. Using an estimation of the projection of the border between every pair of plans we were able to segment all lines used as input and then compute each plan's set of parameters.