How to improve spatial detection position accuracy

leets0429

I have an oak-d-pro-w and I performed calibration based on this guide https://docs.luxonis.com/hardware/platform/depth/calibration/

my reprojection errors are:
rgb: 0.622
left: 0.192
right: 0.203

I used the following script for spatial detection
https://docs.luxonis.com/software-v3/depthai/examples/spatial_detection_network/spatial_detection/

however the position returned by neural network has a higher deviation when the object is further away from the image center. The distance between and camera and the ground is 105cm

My questions are as follow:

Are my reprojection errors low enough? What are the best reprojection errors I can achieve? How can I imporve the calibration
I want to have -+1cm accuracy, is that feasible with my current hardware? What are the position accuracy given a well calibrated camera
Is there anything else affecting the position accuracy

jakaskerl

leets0429

They seem low enough, but you are getting errors of over 10% at sub 1m, so this is not expected. What do your depth images look like?
https://docs.luxonis.com/hardware/platform/depth/depth-accuracy/#Depth%20Accuracy-Stereo%20Depth%20Accuracy-800P%2C%20Wide%20FOV%2C%2075mm%20baseline%20distance%20OAKs
Environment conditions

MTVZ

Good Morning,
me and leets0429 who started this post are working in the same company,
so the following is related to the question from leets above. The depth image and environment relate to the data from him standing at the table above.

Indoor, Artificial Lighting, Checkboard surface for texture.
Distance about 1.5m, Camera is mounted.

jakaskerl

MTVZ
The depth quality is quite bad. This is not due to a calibration issue (mostly), but because the target you are looking at is not textured enough. Pixels from one camera can not be matched with pixels from the other one (or they are matched to the wrong pixels). This produces the cross pattern visible in your depth map - even though the surface is flat.
Turn on the dot projector to improve texture - or use a different target surface.

Thanks,
Jaka

MTVZ

Hello, I turned on the dot projector, but I am not seeing the dots being projected in the images.
Quality also remains unchanged; the reference picture shows how I implemented turning on the IR.
Is that done correctly? I'm using Depthai v3 since we use the NN-Archives module for custom OD.
When using the v2 depthAI I had no problem turning on the dot projector, and I saw dots being projected in the images too.

MTVZ

To visualize what I'm trying to do, we are trying to get positional estimates for object-detected entities.
Visual dots turned on and off, as seen above, do not affect the depth image seen.

jakaskerl

MTVZ
Can you try moving the setIR... call after the pipeline.start()?

Thanks,
Jaka