• OAK-D-PRO-FF very inaccurate depth measurement

Hello,

I'm performing a metrological assessment of a brand new OAK-D-PRO-FF device. This is done by moving a vertical target by means of a translating sled.

I wrote my own code following the tutorial instructions and I double checked the results with those of the examples provided in the Luxonis' repo. Results are quite consistent, mine are a bit better just because I'm using proper filters.

The distance is assessed either averaging or calculating the median value of a 15px side square part of a flat plane of which points are equidistant from the camera. RGB alignment doesn't significantly modify the outcome. The projector is on.

The results that I get are very bad and also very different from all the experiences that I found in the Internet. Please find a summary in the attached file: z difference is higher than 8% at 2 m. Subpixel analysis is active and, if I increase the fractional bits, I get a "less quantized" cloud but equally inaccurate. Using 400_P or 720_P (with decimation 2) resolutions doesn't change the result either.

I don't think that I'm making particular mistakes (if needed I can also attach my code), therefore I'm wondering whether there might be some issues with calibration. Nonetheless, I'd like to get a feedback before recalibrating the system, which is supposed to provide good results even with the factory calibration.

I performed the same procedure (using the same analysis code) with an Intel Realsense D435 and I got errors below 2%, as expected.

Can you please help me?

Thanks

Alessandro

Hi Martin,

thank you for your feedback.

As I wrote in my request, recalibration is for sure something I was considering as the next step. Nevertheless, I was wondering whether I was making some mistakes or missing some important points, as the documentation states that the factory calibration is supposed to give good results, while my errors are big.

Do you think that recalibration is the right way to proceed?

May it happen that some devices leave the factory with such a bad calibration? This wouldn't be a problem, but I need to understand if this happens sometimes or if each camera must be recalibrated (we are considering to use this device for a new product).

Many thanks

Alessandro

Hi Alessandro,

I had the exactly same problems before recalibration, it did a much better result than I expected.

Yesterday I was told that temperature changes can require a new calibration, but I'm not sure how much of a temperature change is needed for require a new calibration.

Br Martin

  • erik replied to this.

    Hi Martin,

    thank you very much for sharing your experience with me.

    I'm going to try. I'll let you know.

    BR

    Alessandro

    Hi Martin,

    I recalibrated the device and the results have dramatically improved.

    Many thanks for your help!

    BR

    Alessandro

      Hi AlessandroB / Robengaard , Thank you for the report, and we apologize for the inconvenience of inaccurate factory calibration - we are soon rolling out factory depth accuracy evaluations which should catch such "bad apples".
      Kind regards, Erik

        Hi erik ,

        thanks for your reply.

        This is a very good news, as final users hardly accept to recalibrate the devices.

        BR

        Alessandro

          AlessandroB Robengaard Could you also share the MxID of your devices? We want to check the flashed factory calibration.

          import depthai as dai
          with dai.Device() as device:
              print(device.getMxId())

          Thanks, Erik

            15 days later

            Hi AlessandroB ,
            Following up, it looks like we don't have logs for this device (we log since May 2022), so unfortunately, we can't check the calibration accuracy. We have since also updated out calibration process (with TVs rather than robot hand), which improved the calibration in general. I do hope such issues won't be encountered in the future anymore.
            Thanks, Erik