Hello ! Looking at the point cloud generated from the depthmap, we noticed it was pretty noisy. [upl-file uuid=c41f68c4-8d8b-4d55-ba8c-91c3b78fa9ad size=4MB]show-pcl2-2024-04-26-114115.mp4[/upl-file] The mug is about 40 centimeters away from the camera. This SR has factory calibration. We tried to recalibrate with another SR and got very similar results. I tried a lot of parameters and filters combinations and this is the best I was able to get [upl-image-preview url=https://discuss.luxonis.com/assets/files/2024-04-26/1714124639-937523-image.png] To be clear, averaging the depth values inside a mask of the mug gives quite a good depth estimate (we are able to grasp objects with our robot using this depth estimate), so the accuracy seems to be OK, but noisy. However, we also want to use point cloud based algorithms, so the cleaner the point cloud the better. Do you think we should be able to get a better point cloud out of this camera ? If so, what can we do to improve it ? Thanks :) Antoine

@"jakaskerl"#1112 Do you have a timeline on fixing this issue ?

OAK-D SR point cloud not great

apirrone

Hello !

Looking at the point cloud generated from the depthmap, we noticed it was pretty noisy.

show-pcl2-2024-04-26-114115.mp4

4MB

The mug is about 40 centimeters away from the camera. This SR has factory calibration. We tried to recalibrate with another SR and got very similar results.

I tried a lot of parameters and filters combinations and this is the best I was able to get

To be clear, averaging the depth values inside a mask of the mug gives quite a good depth estimate (we are able to grasp objects with our robot using this depth estimate), so the accuracy seems to be OK, but noisy.

However, we also want to use point cloud based algorithms, so the cleaner the point cloud the better. Do you think we should be able to get a better point cloud out of this camera ? If so, what can we do to improve it ?

Thanks 🙂

Antoine

jakaskerl

Hi @apirrone
I think it's a problem of the base algorithm that causes the issues. We have seen that the point clouds might not be as good as want and we are working on improving them through the FW. Right now, most you can do (apart from playing with the filters) is to improve the texture of the observed surfaces to increase the depth confidence and precision.

Thanks,
Jaka

apirrone

Ok thanks @jakaskerl !

apirrone

@jakaskerl Do you have a timeline on fixing this issue ?

jakaskerl

Hi @apirrone
No timeline per se, but we have found a bug that was causing issues with rectification so the pointclouds should improve with the new release (still not how we want the end result, but closer). We will be working on improving it further in the upcoming months.

Thanks,
Jaka

apirrone

Hi @jakaskerl

Thanks for the info! When you say "improve with the new release", do you refer to the last release (2.25.1.0) ? or to the next release ?

jakaskerl

Hi @apirrone
In the next release. Probably won't be noticeable on NFOV cameras though. To further improve pcl we need to improve depth which is not such a straightforward process.

Thanks,
Jaka

gdeanrexroth

Has the new version been released yet?

jakaskerl

gdeanrexroth
It has (2.27). Further depth improvements were not made since we are currently focusing on series 4 devices, making sure the launch is successful.

Thanks
Jaka