High-Performance and Tunable Stereo Reconstruction

Author: Pillai

Year: 2016

Notes:

reconstruction with stereo at 120 Hz with a stereo accuracy as input
Semi Global Matching & ELAS = soa in stereo matching
Only fpga or parallel processor based methods can reach rt
DDPTAM: dense mapping from piecewise planar surface with PTAM

Iterative stereo algorithm

First step: construct a piece-wise planar scene with FAST features matched along epipolar lines, and Delaunnay triangulation
It provides a depth prior for the second step: disparity interpolation of every pixel using the plane equation of every triangle
Third step: compute a census (5x5 patch) matching cost of the interpolated disparity
Fourth step: determine costs of region high and low matching confidence and keep the disparity with the lowest cost in the disparity map
Fifth step: support ressampling, region with high matching cost are resampled and the keypoints are mathced via epipolar search
Then the process is performed iteratively

Experiments

Evaluate against line sweep, SGBM and ELAS