Thursday, November 25, 2010

Polling Interface: Sensor Taxonomy

Obviously you wouldn't classify a turkey as a bird, since they can't fly.
What would you classify the Kinect Sensor as?

Your suggestions in the comments, and your poll answers on the sidebar to the right.


Andrés said...

My attempt!
DOF Sensor= Depth Of Field sensor

Julien said...

Hi. "Sensor device" as a family, or even "multiple sensors device", followed by the usage that you decide to make of the Kinect : "as a 3D sensor" or "as a camera". I would not say "3D camera" because it's not really a camera filming in 3D (no stereovision).

j_hodgie said...

I think it is classified as an RGB-D camera (red, green, blue, - depth)

I Heart Robotics said...

DOF Sensor seems easily confused with 3 DoF Sensor, as in degree of freedom. On the other hand I like that it contrasts with ToF, Perhaps there is another XoF that could describe it.

As far as I can tell internally the Kinect is basically a set of stereo camera and or project combinations. The on-board processing obviously helps make this realtime.

I would state that a sensor is a camera if it is best described using projective geometry like the pinhole camera model.

Here is another question
Mono, Stereo, ?, Quadra

I Heart Robotics said...

Some of the current research, most notably from Professor Deiter Fox's group has been using the term RGB-D Camera.

I think the current I Heart Robotics style guide is to call these type of things either '3D Sensors' or 'RGB-D Sensors'

Julien said...

1,2,3,4 : mono, stereo, tri, quadri (from latin).

Examples : tri-filter or triphonic, quadriphonic

(although there are two cameras, I have not yet seen that Kinect uses both togethers)

acho said...

There is plenty of 3D-related sensors, so it would be meaningless to call it "3D sensor".

If I have to pick one, I'd say "RGBD sensor", but Kinect also provides infrared images and audio (maybe even more). Thus, its clearly a multiple sensor device: RGBD+Infrared+Microphone

I Heart Robotics said...

The IR Laser projector and the IR camera form a stereo pair. The math is basically the same as for stereo vision except that with the projector the direction of the vector is reversed.

My concern has been disambiguation. If you say "My algorithm requires a 3D sensor", someone might reasonably expect that a Swiss Ranger 4000, a Velodyne HDL-64E, a Hokuyo URG04-LX mounted on a rotary stage and a Kinect might all work. So there should be a general term to disambiguate the Kinects additional features, which is why RGB-D seems popular. For example the addition of color can improve the performance of ICP as the color space vector can be incorporated in the distance metric. This can be used to avoid local minimas.