Real Time Shape From Silhouette using Cell B.E.
Tuesday, September 30, 2008, 10:30am - Tuesday, September 30, 2008, 12:00pm
The Shape From Silhouette (SFS) algorithm is a well known approach to 3D shape detection from 2D camera images. It is also a performance bottleneck in markerless motion capture technology for use in move production. We have accelerated SFS using Cell B.E., a heterogeneous multicore chip with 200 GFLOPS peek performance. However peek performance requires highly vectorizable algorithms. We parallelize SFS in a novel way. Our approach demonstrates bitwise SIMD parallelism, sequential RAM access, and power of 2 data alignment. Our implementation on Cell B.E. demonstrates near real time performance at high voxel resolutions.