Custom Models for CPAI

[winston_j_allegath]

Where do I extract and place the ipcam-combined model for use with CPAI? Also, how do i select it in Blue Iris?

I am currently running an i7-11700K, 32GB RAM, Coral m.2 TPU, 1TB NVME boot drive, and a 14TB platter for camera storage. I have 9 cameras running on Blue Iris w/substreams.

 

[VideoDad]

Where do I extract and place the ipcam-combined model for use with CPAI? Also, how do i select it in Blue Iris?

I am currently running an i7-11700K, 32GB RAM, Coral m.2 TPU, 1TB NVME boot drive, and a 14TB platter for camera storage. I have 9 cameras running on Blue Iris w/substreams.

It should already be installed with CPAI. First on the AI settings tab, turn on custom models. (You also might as well turn off the default object models).

For each camera, under the AI settings, type "ipcam-combined" as the custom model. Then type the items you want to detect under "To confirm" separated commas. Of those objects, list which are vehicles in the vehicle field.

 

[winston_j_allegath]

Mine does not show that model.

 

[VideoDad]

Have you started one of the Yolo modules?

 

[quinti]

custom models aren´t available for Coral TPU module, only for the CPU/GPU modules. Many people waiting for them for a long time now, me too...

 

[winston_j_allegath]

ok that makes sense. what model do i pick?

 

[VideoDad]

ok that makes sense. what model do i pick?

The YOLOv5 .NET module is pretty good on a CPU. Then you'll have access to the ipcam-combined model and its objects (person, bicycle, car, motorcycle, bus, truck, bird, cat, dog, horse, sheep, cow, bear, deer, rabbit, raccoon, fox, skunk, squirrel, pig)

 

[winston_j_allegath]

i used that for years. i was sick of my cpu spiking to 100%, hence the justification for a TPU

 

[quinti]

ok that makes sense. what model do i pick?

ive tried mobileNet SSD, efficient Det and yolov8, found no differences between them in time or accurancy. Same with size tiny, small or medium. Only with large model times are 10x higher.
In all cases ive more then 10% failed inferences, very bad...

 

[winston_j_allegath]

ive tried mobileNet SSD, efficient Det and yolov8, found no differences between them in time or accurancy. Same with size tiny, small or medium. Only with large model times are 10x higher.
In all cases ive more then 10% failed inferences, very bad...

Bummer. Oh well. I will just wait