This is the source code for my "Surface prediction for a single image of urban scenes" paper. If you use this software or annotated data, please cite the paper:
F. Akhmadeev. Surface prediction for a single image of urban scenes. Computer Vision - ACCV 2014 Workshops, volume 9008 of Lecture Notes in Computer Science, pages 369–382. Springer International Publishing, 2015.
The source code has been tested under OSX 10.9 and 10.10. Pure surface prediction approach should run under linux and windows too.
To start the demo simply run demo_sp in matlab.
The surface prediction approach uses an initial orientation map from geometric reasoning to build final orientations of surfaces. In addition, vanishing points and lines are computed using the code from geometric reasoning by default.
Surface prediction approach can be combined with geometric context.
To build GC you need to run make command in lib/segment/ (segmentation) folder and build mex file from lib/GeometricContext/src/boosting/treevalc.c.
To use this algorithm you need to run make command in lib/VPdetection/ folder. Then, uncomment [vp,f,linesmore] = main(img2);
line in demo_sp.m. Next,
[lines, linesmore] = generate_lines(img2);
[vp, f] = compute_vp(lines, imsize2);
comment those lines.
The source code for vp detection has been modified by me so that it can run under OSX and opencv 2.4.9.
This library is not compatible with windows.
We provide annotated data for Delage et al. dataset and York Urban database. See data/ folder.
The paper was originally presented on the SUAS 2014 workshop.
The project page is available here.