What Are the Best Algorithms for 2D to 3D Scene Extrapolation in Egomotion?

[Superposed_Cat]

:ey all, I'm seeking to write an egomotion program and want to do research into existing algorithms and the math of it, but googling "egomotion algorithms" doesn't seem to turn anything up, is there a more popular name for extrapolating 3d scen from 2d scene? Or can anyone post a link? Any help apreciated,

 

[Fooality]

Do you have to extrapolate from just a 2D image? That puts some limitations on you that you wouldn't have if you used a solution like Kinect, that builds the 3D scene from laser points. In fact the best solutions are generally going to involve projecting light of some form.

As far as getting 3d info from 2d, Android has an interesting app, you can see its output here (if you have a browser with webgl)
https://www.chromeexperiments.com/experiment/android-lens-blur-depth-data-player
http://www.clicktorelease.com/code/depth-player/
That one uses no projection, a single camera (from phone) and refocuses it I believe to get depth data based on which areas are sharp and blurry at each point of focus. If you had a good robot, where its wheels never slip or anything, you could potentially add average in snapshots from other position if the environment is stationary, to get an even better picture.

 

[kroni]

Egomotion determination from 2D picture is call SLAM (simultanéous localisation and mapping), Monocular SLAM. The best class of these algorithm is BUNDLE ADJUSTEMENT which is an non linéar optimisation of reprojection error of feature. You can search for downloadable algorith on MRPT library, OpenSLAM, LSD Slam... I am an old computer vision PHD, if you have question, ask me.

 

[Fooality]

Egomotion determination from 2D picture is call SLAM (simultanéous localisation and mapping), Monocular SLAM. The best class of these algorithm is BUNDLE ADJUSTEMENT which is an non linéar optimisation of reprojection error of feature. You can search for downloadable algorith on MRPT library, OpenSLAM, LSD Slam... I am an old computer vision PHD, if you have question, ask me.

A computer vision PhD? Actually I have a question: Do you know or have any educated guesses about what the spinning sensor on the Boston Dynamics robots is? Video here:

I can make guesses, that its maybe projecting some sort of plane and calculating depth by offset with a camera, but I don't know.

 

[kroni]

Yes, This spinning sensor is a panoramic laser range finder, may be a Riegl, or a Velodyne. It'is composed of 32 or 64 laser verticaly and when the sensor spin it give a point cloud représenting the scene. I work with this kind of sensor in my research.

 

[Fooality]

Yes, This spinning sensor is a panoramic laser range finder, may be a Riegl, or a Velodyne. It'is composed of 32 or 64 laser verticaly and when the sensor spin it give a point cloud représenting the scene. I work with this kind of sensor in my research.

Ah! If I'm not mistaken then, the principle behind it has actually been around for awhile, but it still must be best suited for the job, if they're using it. Fascinating, and thanks for your reply!