OAK-D Pro poor performance compared to OAK-D & Realsense D415

Vinay

For the point cloud of a flat surface OAK-D Pro seems to be doing worse than Realsense D415 and OAK-D.

The mean z value for OAK-D, OAK-D Pro and D415 are 6.74mm, 25.93mm & 10.88 mm respectively, while the standard deviations are 2.63mm, 5.13mm & 2.69. Z=0 represents the flat surface

What steps can I take to have OAK-D pro perform at the same level better than OAK-D.

Data Files : Google drive url

Setup summary:

All the three cameras were approximately 1 meter away from the checkerboard. Their extrinsic calibration were computed using checkerboard detector and the following ros topics:

d415: /camera/infra1/image_rect_raw

oakd_pro & oakd: /oak/left/image_rect

Modification in the default ROS packages

d415:

temporal_filter.enable: True

high_accuracy.json settings

oakd_pro & oakd:

rgbd.yaml: camera:

i_enable_ir: True
i_laser_dot_brightness: 1200
stereo:
i_subpixel: True
i_publish_left_rect: True

OAK-D Pro POV

jakaskerl

Hi @Vinay

Redo camera calibration (https://docs.luxonis.com/projects/hardware/en/latest/pages/guides/calibration/)
i_laser_dot_brightness should be 800 for max luminosity.

Vinay Their extrinsic calibration were computed using checkerboard

That might not be accurate.

Vinay The mean z value for OAK-D, OAK-D Pro and D415

Compute median of the board and discard outliers. Should give a much better result.

Thanks,
Jaka

Vinay

jakaskerl

Hello @jakaskerl , thank you for your prompt response.

I will report the results with re-calibration and 800 laser dot brightness.
What do suspect the possible source of errors in the extrinsic calibration might be? I am using Pinhole camera model, fromCameraInfo method, /oak/left/image_rect image topic (which I think is rectified), and OpenCV checkerboard detection methods (1 , 2). Also the checkerboard is UV printed to correct dimensions.
I plan to compute volumetric estimates of irregularly shaped solids after calibration and so discarding outliers will not work for that. I have been using the median z value from flat surface to modify the extrinsic calibration and it does improve results. However it is the wider spread of z value for OAK-D pro (larger standard deviation) that I am more concerned with.

Thanks,

Vinay

jakaskerl

Hi @Vinay

Vinay What do suspect the possible source of errors in the extrinsic calibration might be?

Your accuracy is limited by the sensor resolution. So you ground truth might be skewed by these small inaccuracies. The rectification can be a source of errors as well.

Vinay However it is the wider spread of z value for OAK-D pro (larger standard deviation) that I am more concerned with.

Take more images if you can and average the values temporally.

Thoughts?

Vinay

@jakaskerl Thank you. Re-calibration improved the mean z distance and standard deviation. However the standard deviation still remains large. Is there some additional method than the calibration script mentioned above.

Camera: D415 Mean: 10.88 Std: 2.69

Camera: OakD Mean: 6.74 Std: 2.63

Camera: OakD_Pro: Factory Mean: 25.93 Std: 5.13

Camera: OakD_Pro: Updated Mean: -2.29 Std: 3.79

Regards

-Vinay

jakaskerl

Hi @Vinay

Try lowering the confidence threshold: depth.setConfidenceThreshold(100)
Apply postprocessing/filtering like median filter.

Thanks,
Jaka