Parallel Visual Hull Computation
Master thesis
Visual Hulls are 3D models of objects obtained by backprojecting in space a set of silhouettes of the objects. They are of great importance in many image based modelling applications where they are either used as an initial model that can be improved or as a full model. Critical issues when computing visual hulls are then twofolds: robustness that ensures a proper 3D model as the output and computation time that allows real time or quasi-real time applications. Most existing approaches do not address both issues and provide partial solutions that are either fast but unprecise, e.g. voxel based approaches, or precise and exact but slow, e.g. current mesh based approaches.
In this work, we propose to address both issues by exploring a new direction where visual hulls are computed by a set of distributed and independant operations. To this purpose, we will exploit the fact that the visual hull is the intersection in space of several generalized cones, i.e. the back projection of a silhouette in space.
Such intersection can thus be obtained by combining two by two intersections of generalized cones in a hierachical way. Challenges are then how to ensure exactness in each operation/intersection as well as how to properly distribute the two by two intersections in order to optimize time complexity.
This work will be conducted in collaboration with the Perception team (computer vision) and the Moais team (parallel computations). We expect results to be implemented and demonstrated on the Grimage platform: a real time multi-cameras acquisition platforms.
References:
A Distributed Approach for Real-Time 3D Modeling J.-S. Franco, C. Menier, E. Boyer and B. Raffin IEEE CVPR Workshop on Real-Time 3D Sensors and their Applications, 2004
The Grimage Platform: A Mixed Reality Environment for Interactions J. Allard, J.-S. Franco, C. Ménier, E. Boyer and B. Raffin International Conference on Computer Vision Systems (ICVS’06)
Start date: 1 March 2009
Contact person: Edmond BOYER
Deadline: 1 November 2008
