SPS-Stereo: Slanted Plane Smoothing Stereo

SPS-Stereo is a dense stereo method employing a slanted plane model. It jointly estimates a superpixel segmentation, boundry labels (such as occlusion boundaries), and a dense depth estimate from a pair of stereo images.

Citation

@inproceedings{Yamaguchi14,
author = {Koichiro Yamaguchi and David McAllester and Raquel Urtasun},
title = {Efficient Joint Segmentation, Occlusion Labeling, Stereo and Flow Estimation},
booktitle = {ECCV},
year = {2014}
}

Building SPS-Stereo

1. Prerequisites
   • opencv
2. Building
   1. type 'cmake .'
   2. type 'make'

Usage of demo code

First, download KITTI stereo/flow dataset from KITTI Vision Benchmark Suite homepage and extract it.

Run SPS-Stereo
> ./spsstereo images.txt

where images.txt looks like: left_image_1.png right_image_1.png .. left_image_N.png right_image_N.png

Outputs

• 000000_10_disp.png
  Disparity image (PNG 16bit grayscale format)
  (Disparity value) = (Pixel value)/256.0

• 000000_10L_seg.png
  Segmentation image (PNG 16bit grayscale format)
  (Segment ID) = (Pixel value)

• 000000_10L_plane.png
  Estimated disparity planes
  the number of lines = the number of segments
  Each line includes parameters of disparity plane of a segment: (A_i, B_i, C_i)

• 000000_10L_label.txt
  Boundary labeling result
  the number of lines = the number of boundaries
  Each line includes three entries: SegmentID1 SegmentID2 boundary_label
boundary_label: 0 (Occlusion, SegmentID1 is front), 1 (Occlusion, SegmentID2 is front), 2 (Hinge), 3 (Coplanar)

• 000000_10L_boundary.png
  Visualization of segmentation result
  Boundary color means a type of relationship between neighboring segments: Red/Blue-Occluion (Red side is front), Green-Hinge, Gray- Coplaner

License

SPS-Stereo is licensed under the GNU General Public License Version 3 (GPLv3), see http://www.gnu.org/licenses/gpl.html.