Code for the paper "Unsupervised Generation of a Viewpoint Annotated Car Dataset from Videos", ICCV 2015
The code is provided for research purposes only. Any commercial use is prohibited. If you are interested in a commercial use, please contact the copyright holder.
Please cite the following paper if you use this code or its parts in your research:
@InProceedings{SB15,
author = "N. Sedaghat and T. Brox",
title = "Unsupervised Generation of a Viewpoint Annotated Car Dataset from Videos",
booktitle = "IEEE International Conference on Computer Vision (ICCV)",
year = "2015",
url = "http://lmb.informatik.uni-freiburg.de//Publications/2015/SB15"
}
Please report bugs and problems to Nima Sedaghat ( nima@cs.uni-freiburg.de )
External software/libraries -- not included:
- VisualSFM (http://ccwu.me/vsfm/) + SiftGPU (http://www.cs.unc.edu/~ccwu/siftgpu/)
- SSD (http://mesh.brown.edu/ssd/)
External software/libraries -- included in the 'extern' directory:
- Compute mesh normals (http://de.mathworks.com/matlabcentral/fileexchange/29585-compute-mesh-normals)
- ransac (http://www.peterkovesi.com/matlabfns/index.html)
only these files are necessary: ransac.m fitplane.m ransacfitplane.m iscolinear.m
Download input car videos from here -- videos of 52 static cars, covering almost the whole 360 degrees around the car. Download resulting dataset from here -- viewpoint- and bounding box-annotated car images.
Some external components show random behaviour, which result in non-deterministic results that cannot be exactly reproduced across multiple runs. The closest you can get to a deterministic behaviour is by:
- setting 'settings.deterministic = 1' in the file setup.m
- setting 'param_deterministic_behaviour 1' in the file [VisualSFM Path]/bin/nv.ini
However, the results can still vary noticeably -- unless you use another SFM package, where you can control its behaviour completely.