Object custom detector returns NaN coordinates when detection is in Y < 0 region

[bdeoli]

Hi all,

I’m using the object custom detector with a model I trained myself, and I’m trying to extract the real-world 3D coordinates of the detections. I haven’t changed anything in the code — I’m simply loading my custom model.

Here’s what I’ve observed:

• When the detected object is in Y > 0, the detector finds it, tracks it, and returns valid coordinates. Once tracking starts, I can move the object anywhere in the image, and it continues to work correctly.

• When the object is in Y < 0, it gets detected, but it’s not tracked, and the coordinates are returned as NaN. If I then move the object into Y > 0, tracking starts and the coordinates are valid again.

I’d like to be able to obtain valid coordinates across the entire image, not just in the Y > 0 region.

Has anyone experienced this issue or knows what might be causing it? Is there a parameter or setting I need to adjust to ensure consistent coordinate extraction?

I’m using Right-handed with Y pointing up and I already tried to disable the track.

[bdeoli]

here is one example what happen

[BenjaminV]

Is the camera looking downward here? If so, could you try the same test with the camera facing straight ahead (not tilted downward) but keeping everything else the same?

[bdeoli]

Yes,
Now it’s working after turn the camera. Thanks.
Could you explain why?

[BenjaminV]

Not yet. I’d need to reproduce it on my side, but my intuition told me it was the source of the issue

Thanks for the feedback. I’ll log the issue internally, and it should be fixed in a future patch.