Search code examples
opencvcamerawebcampose-estimation

RGB-Webcam for 3D pose estimation

I don't know whether this is the right place to ask this question but here we go. I'm currently implementing a project for 3D pose estimation and with VOXELPOSE. (based on OpenCV)

Therefore, I want to set up at least two cameras to optimize the 3D triangulation. I thought about buying two identical webcams below 50 €. (low on budget)

Now the question(s): Is there anything in particular I should consider when buying the webcams? Does anyone maybe have any webcam recommendation?

My criteria for now would be that it has a view angle of at least 100 degree, 1080p & 30fps. Do you agree with my criteria? Is there anything else in your opinion?

Thanks for your suggestions.

Solution

  • This is not really a question for SO, but briefly:

    There is no general answer, because one does not do "3D pose estimation" in a vacuum. Before looking at camera spec sheets you need to make explicit the kind of scene, or scenes, you will be looking at and, if the camera is moving, what kind of motions are expected. Some specific questions you need to look at:

    • Indoor or outdoor scenes?
    • Static subjects, or a combination of static and dynamic scene elements?
    • Do you control the scene (e.g., can you place objects of known geometry in it)?
    • Do you control the lighting (location, direction, intensity, timing)?
    • If you do not control the lighting, what variation in illumination can you tolerate?
    • What level of accuracy are you targeting? Are you OK with being 1m or 1deg off the truth, or will someone die / lose a ton of money if you are?

    If you are mostly interested in software, I strongly recommend skipping the hardware altogether, at least initially, and use one of many available open datasets for pose estimation. For example, EUROC-MAV.