View my account

Comments

6 comments

  • MartyG

    The depth accuracy of the D400 series cameras is less than one percent of the distance from the object. So if the camera is 1m from the object, the expected accuracy is between 2.5mm to 5mm.

    The new RealSense L515 lidar camera is also stated in its official pre-release FAQ to have a depth error average of < 5 mm at 1 m range.

    In real world conditions though, accuracy is more complex than that and a range of factors such as camera depth units, resolution and environmental lighting may factor into it.  The D435 also has around double the 'RMS Error' (depth noise that increases linearly over distance) compared to the D415 model due to its hardware configuration.  These concepts are explained in Intel's camera tuning guide.

    https://www.intel.com/content/www/us/en/support/articles/000027833/emerging-technologies/intel-realsense-technology.html 

    The guide contains a formula for calculating RMS Error.

    You may also find useful a white paper that Intel published on depth testing methodology.

    https://www.intel.com/content/www/us/en/support/articles/000026982/emerging-technologies/intel-realsense-technology.html 

    0
    Comment actions Permalink
  • Masahiro Ogawa

    Thank you for your reply. So based on the 2nd document page 7, the ground truth is measured by "laser rangefinder or tape measure" for "a flat surface (target) aligned parallel to the depth camera". Is this correct?

    0
    Comment actions Permalink
  • MartyG

    It looks as though the document is suggesting that a laser rangefinder or tape measure are methods by which people can calculate ground truth, but the camera itself does not do the full ground truth calculation.  

    In Intel's instructions for the RealSense SDK's Depth Quality Tool program, they suggest enabling the 'Ground Truth' option in the tool's user interface and then modifying it with the values obtained from external tools (such as the ones mentioned above).

    https://github.com/IntelRealSense/librealsense/tree/master/tools/depth-quality#quick-start 

    Ground truth is also discussed later in the document.

    https://github.com/IntelRealSense/librealsense/tree/master/tools/depth-quality#z-accuracy 

    0
    Comment actions Permalink
  • Masahiro Ogawa

    Thank you for your quick reply. I didn't notice Intel already provide depth accuracy evaluation tool with GUI. That would be really helpful. 

    I understand Intel recommends to use laser rangefinder or tape measure as a ground truth.

    So just for confirmation, even for getting the nominal value(<1%), Intel used those. Correct?

    0
    Comment actions Permalink
  • MartyG

    I am not on the Intel developer or engineer team, so a member of those teams will be better able to tell you about how RealSense handles ground truth.   You can enquire to them by visiting the RealSense GitHub forum at the link below and clicking on the New Issue button to post a question.

    https://github.com/IntelRealSense/librealsense/issues 

    I do know that for an advanced project where up to 20 RealSense cameras were used to stitch together point clouds into a single large cloud, the CONIX Research Center at Carnegie Mellon calibrated using an expensive precision laser surveying device called a Theodolite.

    https://github.com/conix-center/pointcloud_stitching 

    0
    Comment actions Permalink
  • Masahiro Ogawa

    Great! Thank you Marty for your very useful information.  

    0
    Comment actions Permalink

Please sign in to leave a comment.