Stairs: Testing Stereo Matching Algorithms at Occluding Boundaries
Abstract
When occluding contour and texture discontinuity coincide, area-based
matching algorithms give wrong disparity estimates near the
contour. The contour `leaks' into one of the regions neighboring
it. (Usually, the strong-textured region leaks to the weaker-textured
region.) We have developed a method for ranking area-based matching
algorithms according to the relative area of the artifacts. Evaluation
method is described and results on some algorithms are presented. Our
data and ground-truth are available from this page.
The Test Target
|
|
|
side view
| front view
| back view
|
The staircase-like target (left) is designed to combine all
possible texture contrast changes across left or right occlusion
boundary (center) or across no depth change (right). All textures are
random binary Gaussian noise of different mean gray-value and
contrast. The texture pattern is the same on the target's flat back
so that the influence on the matching algorithm of the texture
boundaries alone can be studied in a control experiment.
PostScript textures for the three panels are available:
front,
middle,
rear.
Rectified Images
Images were taken by calibrated Sony XC-77CE TV cameras, were
rectified and sub-sampled to 256×256 pixels.
|
|
front: left image
| front: right image
|
download: [
GIF,
PM,
PGM]
| download: [
GIF,
PM,
PGM]
|
|
|
back: left image
| back: right image
|
download: [
GIF,
PM,
PGM]
| download: [
GIF,
PM,
PGM]
|
Ground-Truth Data
Ground-truth disparity map and occlusion maps were obtained from a
disparity map recovered by NCC area-based matching with sub-pixel
accuracy. Occlusion boundaries were identified manually with pixel
precision. Within each flat region of the disparity map a plane was
fit to the map, the fit was sampled (see below, left) and occluded
regions were automatically identified (see below, right).
|
|
disparity map (pseudocolors)
| left occlusion map
|
download: [ASCII]
| download: [ASCII]
|
|
disparity map
|
---|
download: [ASCII]
|
Both datafiles are plain ASCII files with 256 rows of 256 IEEE-format
double integers. You may download them directly to your Matlab using
load
. Ground-truth sub-pixel disparity map pixels are
NaNs where the disparity is not defined (the corresponding matrix
entry is `NaN'). Ground-truth left occlusion map pixels are 1's in
areas that are occluded in the right image and 0's elsewhere. The
disparity and the occlusion maps are not defined on the background
since the background is not considered a part of the ground-truth.
Contour Artifacts Evaluation Method
References
- R. Sara and R. Bajcsy. On Occluding Contour Artifacts in
Stereo Vision Proc. IEEE Conf. CVPR '97,
pp. 852-857. [PostScript, 600kB,
gnu-zipped]
Go to: [Radim's home page |
CMP]
Radim Sara
Last modified: Fri May 7 15:34:00 MET DST 1999