Exploitation of Interframe Redundancy for Real-Time Volumetric Reconstruction of Arbitrary Shapes

Scalable systems reconstructing arbitrary shapes in real time need a cluster of computers and an efficient strategy to share information. Optimizing the performances of the system requires considering how to spread the complexity over the nodes of the cluster. It means taking into account the kind of information to share between nodes, how to transmit it, and how to merge it given the available bandwidth and the impact of those steps on performances. We present a distributed and scalable volumetric architecture based on an efficient exploitation of interframe redundancy and on an efficient merging of partial models. The architecture is composed of acquisition nodes reconstructing partial models from multiple views and of a master node merging partial models. The master node updates local copies of the partial models with nonredundant information from the acquisition nodes. Then it merges the partial models to produce the volumetric description of the scene. The test of this architecture with a single feature coding visibility, occupancy, and subdivision of space proves its efficiency. The chosen feature allows each camera to see only part of the volume of interest. We show real-time results on a cluster of usual PC platforms