The KITTI Vision Benchmark Suite

You can get a copy of the paper here.

Appearance-aligned stereo warps one stereo image according to the expected road geometry for reducing the perspective appearance change between the left and right stereo image. The resulting disparity map is combined with the disparity map of stereo computation without warped images by means of a pixel-wise correlation value comparison. This way warping model coherent pixels are mainly drawn from the aligned disparity map and model incoherent pixels are drawn from the normal disparity map.

For the actual stereo computation a standard block-matching stereo with SNCC cost function is used. The images are pre-processed with a Gaussian filter and resized to half the image size. After the stereo processing the disparity maps are rescaled to full image size. Furthermore, standard post-processed is applied, including a left-right check, error island removal (region grouping), hole-filling and median filtering.

Timings: block-matching SNCC stereo (38ms), LR-check (18ms), region growing (2ms), hole-filling (2ms), median filter (23ms)

Runtime includes some IPP routines but no multi-core or GPU usage.

linear warping model: x2 = x + a*y + n
a=0.25, n=-21.6
SNCC (first stage 3x3, second stage 9x9)
sigma_gauss = 1.0
disp_range = 64
min correlation threshold = 0.3
region growing threshold = 2.5 disparity
min region size = 200 pixel
median filter = 1x9 and 9x1

@InProceedings{Einecke2013IV,
author = {N. Einecke and J. Eggert},
title = {Stereo Image Warping for Improved Depth Estimation of Road Surfaces},
booktitle = {IV},
pages = {89-194},
year = 2013
}