Python tutorial that uses ORB feature from OpenCV to initialize RGBDOdometry

[syncle]

This issues is derived from #24. Lower priority.
Need not to be done before the first launch.

[syncle]

Another idea is providing python example code using OpenCV python interface. We have plan to include OpenCV python + Open3D OdometryRGBD example for #51. In this manner, Open3D does not need to be directly linked to heavy OpenCV.

[syncle]

Working on this issue. But there are few concerns

• Even recent OpenCV distribution only supports Python2.7 (tell me if I am wrong) that might limit our python example.
• We need to include RANSAC+5pt algorithm into Open3D if we go with sparse features. However, the new module will dependent on external feature detector/descriptors. The new module cannot work alone and there is no module like this in Open3D yet.

[qianyizh]

1. OpenCV should support both Python 2.7 and 3.x.
   Try the following command if you are using Anaconda.
   Actually, I highly recommend that we only test python environment (both 2.7 and 3.x though) of Anaconda. It configures almost everything we need.

conda install -c conda-forge opencv

2. Yes, we need to implement RANSAC and 5pt algorithm. I suppose they should also go into the Odometry module.

3. Please don't touch image.h/cpp, RGBDImage.h/cpp yet. I have a big commit that has massive changes to these files. Will be able to push in tonight. You will do the review for me, 👍

[qianyizh]

Actually earlier we had a discussion with the OpenCV core development team, and the main recommendation I gave them was to put OpenCV into Anaconda. (Not in the forge. Forge is like the community build support.) I think they will be working on it. So hopefully in the future, cv will be included natively in Anaconda. We just test this environment.

Btw: Anaconda is also the easiest way for setting up TensorFlow. I think everything is converging into it. It won't be a bad choice if we just stick with it.

[syncle]

Some technical issue.

Please check current implementation of 5pt algorithm. (Let me know if you have not privileged to see the code).

Checkout getCoeffMat function. There is hundreds of hard-coded linear algebra lines that will look not beautiful if they are included in Open3D. The function is not intuitive compared to beautiful Jacobians we have established by chain rule.

It also depends on complex OpenCV function. To name a few solvePoly, that solves high-order polynomial equations. Open3D does not have mathmatical toolbox for solving such problems. Transferring this heavy OpenCV dependent function will add quite amount of work to our workload.

[qianyizh]

Wow, for real! Where did you get these equations?

An alternative approach is to replace the 5-pt algorithm with RANSAC. We do RANSAC on ORB features, and compute transformation based on points converted from depth pixels. Does this make sense?

[syncle]

I decided to go with OpenCV python binding as it provides ORB feature + 5pt algorithm. This is handy and we don't have to merge complex 5pt algorithm in Open3D.

I finished writing a code for building fragment. There will be PR regarding this module soon.

[syncle]

This issue handled with #74