now ELAS library is used by default to compute horizontal disparity map from rectified images

[GiuliaP]

• added ELAS library's source code (lib/elas_src, lib/elas_src_omp).
• modified the lib/CMakelists.txt in order to compile ELAS (also OpenMP accelerated version on UNIX).
• added a class wrapper for ELAS (elasWrapper.cpp, elasWrapper.h) to deal with OpenCV's cv::Mat.
• added 2 functions to the stereoCamera class to initialize and dispose the ELAS wrapper.
• the SFM module initializes the ELAS wrapper by default: to use the previous OpenCV's SGMB you need to pass to the SFM module the --use_sgbm parameter.
• if the ELAS wrapper has been initialized, stereoCamera uses it to compute the disparity map (see stereoCamera::computeDisparity function).
• added 2 output ports to the SFM module to read the rectified images.

[pattacini]

Thanks @GiuliaP could you provide doxygen documentation about the modifications you applied (e.g. the --use_sgbm option, the two new ports ...)? I will be able then to recompile the documentation.

@Tobias-Fischer here's the PR; feel free to participate/test/contribute.

[GiuliaP]

Sure, thank you!

[Tobias-Fischer]

Hi @GiuliaP,

Thanks for the nice contribution! I quickly had a look at the source code, and except for some smaller comments it looks all good (see above & in https://github.com/GiuliaP/stereo-vision/commit/7b1eb3889396a2b2719dc93178be3a685d1fc05e, sorry for being picky ;) ). I'll give it a try on our robot in the next couple of days and let you know how the accuracy is.

Thanks again and best wishes, Tobias

[pattacini]

Thanks @GiuliaP :+1:

[GiuliaP]

@Tobias-Fischer @pattacini Following your observations about the different settings of parameters, I'm modifying the code, providing the possibility of tuning some more parameters. I'll also do some other tests on our robot, trying to define the best configuration. I'll let you know when everything is ready to be tested!

[pattacini]

yes, the idea is to test the final settings along with the depth2kin experiment, to see whether the finger will be always detected.

[Tobias-Fischer]

Hi @GiuliaP: Thanks a lot for your effort! Please let me know once you tuned the parameter set on your robot, and I'll give it a try on ours. Best, Tobias

[Tariq-Abuhashim]

Hi;

Please note that unlike opencv, which performs dense texture matching, ELAS estimates the dense disparity using MRFs (hence it is faster). Therefore, these parameters are going to be scene/application dependent based on how much smoothness is assumed.

Best regards; Tariq Abuhashim

On Thu, Jun 11, 2015 at 10:38 AM, Tobias Fischer notifications@github.com wrote:

Hi @GiuliaP https://github.com/GiuliaP: Thanks a lot for your effort! Please let me know once you tuned the parameter set on your robot, and I'll give it a try on ours. Best, Tobias

— Reply to this email directly or view it on GitHub https://github.com/robotology/stereo-vision/pull/1#issuecomment-111042954 .

[GiuliaP]

@pattacini @Tariq-Abuhashim @Tobias-Fischer I have pushed another commit where I have:

• fixed a bug in the OpenMP accelerated version of LIBELAS (which I have reported to the authors)
• changed the way LIBELAS parameters can be set through the ResourceFinder of the SFM module, updating the related documentation in SFM.h
• now the default LBELAS setting is ROBOTICS, however the ipol_gap_width parameter is set to 40 as in the previous version
• unfortunately we have some troubles with the iCub here in Genoa therefore I haven't tested this version yet, @pattacini suggested me to push it anyway to see whether you can give it a try before us
• therefore I haven't tested the depth2kin experiment yet (for which I suspect that it might be necessary to reduce the ipol_gap_width parameter, in order to detect the finger in a stable way)

[pattacini]

Preliminary tests run with depth2kin along with default parameters of SFM showed a good segmentation!