Confidently Stable Monotonic Matching for Standard Stereo Pairs

The examples shown here demonstrate:
  1. The disparity maps are artifact-free even when the stereo-baseline is quite wide.
  2. There is a small number of mismatches:
  3. even in ambiguous areas (like in the west wing of the Pentagon building),
  4. and untextured areas (like in the underilluminated grass in the Birch Tree set or the dark window in the Shrub set),
  5. at the cost that there are some holes (in black).

In all cases 5x5 matching window and normalized cross-correlation was used to measure match similarity. Confidence interval parameters were set to alpha = 20, beta = 0.03. CPU time was measured on PIII 750MHz under Matlab.

Disparity is encoded by color: large disparities (foreground) are red, small disparities (background) are dark blue. Holes are black. Get the ChromaDepthTM 3-D Glasses from Chromatek to view the disparity maps as 3D objects.

The Shrub set (shrub-3 and shrub-21)
image size 480 x 512, disparity range 10 .. 30, cpu time 6.6s
The Parking Meter set (pm-2 and pm-14)
image size 480 x 512, disparity range 0 .. 30, cpu time 8.5s
The Pentagon set
image size 512 x 512, disparity range -10 .. 10, cpu time 7.8s
The EPI Tree set (epi-32 and epi-16)
image size 233 x 256, disparity range 0 .. 35 (ordering violated), cpu time 1.7s
The Birch Tree set
image size 484 x 640, disparity range 0 .. 60 (ordering violated), cpu time 16.3s
The Map set
image size 216 x 284, disparity range 0 .. 29, cpu time 1.6s
The Sawtooth set
image size 380 x 434, disparity range 0 .. 19, cpu time 4.3s
The Venus set
image size 383 x 434, disparity range 0 .. 19, cpu time 4.2s
The Tsukuba set
image size 288 x 384, disparity range 4 .. 15, cpu time 2.3s

Images courtesy of Carnegie Mellon University, SRI International, University of Tsukuba, and Middlebury College.

Radim Sara
Last modified: Sat Dec 1 16:58:57 CET 2001