@theDesertMoon looks like the limit is indeed 60. Not sure where it comes from.
Limitation in object tracking for oak-1
- Edited
GergelySzabolcs Ah, that sucks.
Does the 'theoretical' on this page refer to another method of object tracking that the OAK camera's have. further down the pipeline?
[https://docs.luxonis.com/projects/api/en/latest/components/nodes/object_tracker/]
Maximum number of tracked objects
SHORT_TERM_KCF can track up to 60 objects at once, while all other trackers can (theoretically) track up to 1000 objects at once.
Even though it crashes the camera/pipeline with a critical error it did momentarily track up to 91 items in a frame and I can see tracklet ID's in the 60's, 70's and 80's.
Is there some resource on the host the camera is looking for? It seems like the error mentions "ShaveHandler". Is there an issue with compiling the model with too many/too few shaves?
Thanks for any information you can provide.
- Deon T
theDesertMoon
It's an issue with the object tracker implementation in firmware.
GergelySzabolcs
Okay.
οΈ
Any chance of it being addressed in a future firmware update or is it a closed issue until some future series re-implementation?
theDesertMoon
What's your volume?
- Edited
theDesertMoon in general, we don't have nearly enough engineers to fix all non-critical issues (such as this one) in firmware. That said, if it's a large opportunity, we could do it for an NRE cost, or even for free if it's a large enough opportunity. Thoughts?
Thanks, Erik
GergelySzabolcs erik Sorry, off on the weekends.
We're doing low volume at the moment for PoC purposes across the line, a few OAK-1-Lites, OAK-D-Lite and now we've moved to your POE offerings. We're envisioning fitting machines with 3-6 PoE cameras each, so eventually we're anticipating ordering in larger quantities of around 100 or so cameraβs a year.
My interest lies mainly in the fact that object tracking makes a more accurate speed determination possible since there's no guarantee the cameras will identify the objects in an identical order each time.
Gergely's description of it as "...an issue...in the firmware" made me curious if its a good enough for now or if its being worked also.
I would be interested if there's a quote for the work required that I can present to my manager.
Thanks,
-Deon T
erik I have another question: I see back in May on Discord you asked Luxonis-Szabi a question similar to mine and he replied not at 30 fps. Is there a FPS scaling chart where I can increase the number of tracked objects based on how much I reduce the FPS by?
Hi theDesertMoon , very likely not, but if you do create one feel free to share, so we can add it on docs
- Edited
erik
Ha, I think Szabi was just joking.
camRgb.setFps(5)
# testing MobileNet DetectionNetwork
detectionNetwork.setBlobPath(nnPath)
detectionNetwork.setConfidenceThreshold(0.50)
detectionNetwork.input.setBlocking(False)
# possible tracking types: ZERO_TERM_COLOR_HISTOGRAM, ZERO_TERM_IMAGELESS, SHORT_TERM_IMAGELESS, SHORT_TERM_KCF
objectTracker.setTrackerType(dai.TrackerType.ZERO_TERM_COLOR_HISTOGRAM)
objectTracker.setMaxObjectsToTrack(61)
Even down to an unusable 5 fps, even trying to get one more tracklet fails.
I'm not sure how I missed your reply from last year though about other Object Tracker functionality:
sorry about the inconvenience, this shouldn't happen, but we, unfortunately, can't do anything about it. Object tracker logic is proprietary from Intel and is compiled so we can't change/fix anything.
I'm am still interested in either of the options you mentioned earlier if something has changed however.
Thank you,
-Deon T
I looked into it and doesn't look like an easy-to-fix BUG.
Will look more this weekend if time permits. Will keep you updated.
GergelySzabolcs Thank you, sir!