The KITTI Vision Benchmark Suite

New 2013 libviso2 version with "Active Matching": This are the results of an updated version that uses the active feature match search strategy by leveraging the motion estimate from the last frame. This reduces the problem of ambiguities and improves performance especially in challenging scenarios such as on highways.

LIBVISO2 is a very simple library for efficiently
computing incremental stereo visual odometry: It
matches features in the previous frame (using a
fast multi-stage method), projects them into 3D
and minimizes the reprojection error wrt. the
current left and right image. Outlier rejection is
performed by means of RANSAC. A monocular version
is also implemented, which estimates the
fundamental matrix based on sparse optical flow
features. This library uses SSE3 SIMD instructions
for speeding-up feature matching. A C++ and MATLAB
demo file exists.

default parameters, except for:
nms_n = 3
half_resolution = 0

@INPROCEEDINGS{Geiger2011IV,
author = {Andreas Geiger and Julius Ziegler and
Christoph Stiller},
title = {StereoScan: Dense 3d Reconstruction in
Real-time},
booktitle = {IV},
year = {2011},
month = {June},
address = {Baden-Baden, Germany}
}