Ver 6.0.9.6 Built-In AI - using Delivery or USPS.onnx custom models failing to Detect usps

[MikeLud1]

I was able to capture 2 instances with the new AOI implemented. I really like the way the AOI limits the viewing area making the objects of interest appear to be larger. Thank you for pointing that out to me.
Unfortunately I am not sure it helped with AI detection much? I did get one false fire of a truck taillight as usps. Not sure why it then failed to present any detections for the actual usps truck that "appears" rather clear in view. ( both attached).
Since the AI model resizes to 640x640 I wonder if I should enlarge my AOI to be at least 640 high rather than the 320 it is at now. While I am not interested in the additional area might it possibly help the AI?

I recently added a new camera that also has a view of the road. I am going to add a clone with AOI to that camera to see if it performs any better.
Thanks for all the input.View attachment 246716View attachment 246717

Try having two clones of the original camera with a 640x640 AOI, the width is still 1520

 

[ELA]

Thanks Mike,
Will give that a try. Will also have a second master camera and clone to try on Monday. Will report back then.

 

[Tinman]

I finally had the USPS truck go by. The delivery model worked running yolov26s. BI 6.0.9.8



Here are my settings:

 

[ELA]

Hello TinMan,
Thank you for testing the usps detection.
I am setup to test (3) 640x640 AOI clones on Monday when usps next delivers using the existing v8 model.

I will then try the yolov26s model after that. That looks very promising.
I read your posts on testing yolov26s and see that you posted the models there. I will download those and try it afterwards.

I was able to get a good confidence detection today on an amazon delivery. However I also got several false detections for usps (it seems to like truck tailights.).

 

[Tinman]

More good results today.

 

[ELA]

I got one good and one minimal detection today using the 640x640 AOI views. Thank you Mike!

The newest Camera ("Mailcam3" is different from any previous tests) has a better view and produced the best result.
The original camera I had used during BI5(deepstack days) "MailCam" still seems to struggle more for some reason.
I have included two captures from that original cam. One with a minimal positive result. The other from the same camera
but with a closer view - but still fails to see anything. I know having the door open makes it harder to resolve.

I will test both for a while but it looks like the new camera view will win.

A small detail: For some reason I was unable to get an exact 640x640 AOI on the "Mailcam". I tried several times.
It would show 640x640 as I dragged the defining box. Each time once saved it showed 624x640? This camera has a 90deg
rotation. Does that possibly affect things for that reason?

 

[ELA]

More good results today.

View attachment 246782View attachment 246783

Thanks Tinman,
I downloaded the yolov26s file and will try that in a couple of days after testing with success using 640x640 AOIs.
From what I could understand; you use the yolov26s file as the primary and the delivery model from the standard model package, as an "additional model"?

Unless these models interact in some way, would having the yolov26s as primary help the delivery model work any better?

I will try the yolov26 in hopes of gaining some speed for the primary model. My computer is a real slug compared to yours.
This newer BI6 does strain my computer more than BI5 did and I hope to avoid needing to get a much improved graphics card.
I am finding that the built-in Intel 630 performs better than my very old GeForce1030 card.