I'm trying to distinguish by code between a real 3D object and its picture - a 2D object, based on depth information. However the picture (= the 2D image) might be captured from noisy background, so I need something more sophisticated than just having diversified depth information; also, the depth camera's output itself might be noisy and there isn't a straight-forward correlation between the model shape and the depth information.
For example: I have two pairs of color-depth images. One: color/depth pictures of a doll, second: color/depth pictures of a pictured doll. I want to create a flow that for pair no. 1 generates a "yes" output, and for pair no. 2 a "no" output.

Any advice - academical sources as implementation ideas - are very welcome.

  • $\begingroup$ In the second depth image you should find a plane (of the picture) while in the first image it is a more complex object. The whole plane appears with a constant color if it is parallel to the camera plane. Otherwise its color varies quite like a linear gradient. If this task has no other requirements then you can generate the yes/no answer by checking that plane in the image. Please give more details. By the way, what's the purpose of this task? $\endgroup$ – Bálint Fodor Feb 15 '15 at 18:02
  • $\begingroup$ @BálintFodor: Thanks, this is indeed what the current solution is based on - find the planes in the depth image and uses some threshold on the number and the size of the planes. However the depth information doesn't have that high quality, and there are many false positives and false negatives. The purpose of this task is to detect an object in a picture, but I want to avoid the case when the user just shows the pre-pictured object to the camera instead of showing the real 3D model (as the detection is based on the RGB only). $\endgroup$ – rursw1 Feb 16 '15 at 9:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.