I’m trying to detect holes on a calibration board, however the background wall is “filling” in the holes and being treated as part of the board.
Upper left is the depth camera output, bottom left is the edge detection and the right is the pcl generated by the calibration code.
What can I do to improve the hole detection in this condition?
The holes are empty when used with a simulated zed camera.
can you record and share an SVO in the same conditions so we can test it?
What settings are you using? Have you tried to reduce the depth confidence threshold?
SVO file: https://file.io/FmkC4XEz6jOX
I tried hanging a black cloth behind the board to try and make the distinction between the hole and the “wall” with still no luck. Reducing the depth confidence threshold did show some improvements but it only detected 3 out of the 4 holes at best.
Have also to move the board’s distance from the wall (both closer and further) and the camera’s position and angle without success.
I shared your SVO and your information with the R&D team, we will debug the behavior with your data and we will provide a fix with one of the next patches of the SDK.
What I can say is that it’s not easy for a pure visual stereo matcher algorithm to understand if the holes are drawn on the white plane or if they are part of the background.
That’s not even easy for the human eye because the holes do not have enough visual features to be matched.
A texturized background would be easier to be detected correctly.
I will try to get a textured background to test with the board.