Hello. I have been working with several OAK-D cameras (OAK-D POE, OAK-D S2 POE, OAK-D Pro POE) and have noticed there is a significant amount of absolute depth error that seems to greatly exceed the error quoted here.
I have performed an experiment in which I am obtaining the pose of a ChArUco board (using traditional OpenCV methods) and publishing this to tf2 of ROS 2. I'm then comparing this pose to the point cloud from the camera using RViz. In the images below, the ChArUco board is just under 2m from the camera and the short side of the board is 210 mm long. At this distance, the quoted error of 1-2% would mean the depth accuracy would be wrong by 20-40 mm. However, judging from the size of the board and the error as seen in RViz, it looks closer to 7-8% error. If it will be helpful, I can try to get an actual measurement of the error.
I am using the pointcloud.launch.py launch file which is using subpixel mode.
Left mono image:
Right mono image:
