60 likes | 152 Views
REMINGTON GONG BENJAMIN HARRIS IURI PRILEPOV. Guess the Depth. WHAT WE HAVE DONE. Created a profiler to sample current performance of feature tracking Useful for deciding which of next stages are worth implementing on GPU Started the eight-point algorithm implementation
E N D
REMINGTON GONG BENJAMIN HARRIS IURI PRILEPOV Guess the Depth
WHAT WE HAVE DONE... • Created a profiler to sample current performance of feature tracking • Useful for deciding which of next stages are worth implementing on GPU • Started the eight-point algorithm implementation • Researched SVD-based methods for recovering rotation and translation between correspondence pairs
FEATURE TRACKING PERFORMANCE • Setup: • Video - 640x320 @ 24Hz • Laptop - Intel Core 2 T7500, NVIDIA 8600M GT Note: Average times do not sum to total average time since some phases (i.e. new feature selection) are not always computed.
Performance Observations • Feature tracking on GPU is fast, but GPU memory copy-back during feature selection hurts (5 ms) • Convolution is expensive on GPU due to numerous texture fetches despite trivial parallelism • Lots of room for improvement, but currently capable of 35Hz • Several stages still to be implemented
Eight-Point Algorithm • Recover 3D coordinates from a set of point correspondences between two frames • Least-squares approximation of the Essential Matrix (camera movement) • SVD – Jama linear algebra package • Normalization of input coordinates
OUR ROAD MAP... • Structure From Motion • Complete eight-point algorithm implementation with the normalization of input coordinates • Mesh Generation • Delaunay refinement mesh generation • Texture Mapping