Just switched from CPU to TPU - cpai 2.8.0, BI

[scoob8000]

Running blue iris with default ai settings in config. Cams analyze a few images per trigger only looking for a handful of objects (person,dog,cat, car, etc)

Since switching to TPU (TF lite) processing is waaay faster but accuracy has dropped. I'm getting a lot of fairly confident (60+%) detections that are wrong. Inanimate objects in the yard as person, dog as a bird, a deer last night as a rabbit. Most camd are around 50% to confirm.

I'd occasionally get incorrect ones with CPU, but nothing like this.

Is that the nature of the beast with the TPU model? I'm not sure what size model it's using either. Just the default when I installed the coral module.

Open to suggestions.

 

[wittaj]

Yep that is pretty common. Many that tried TPU went back. Faster isn't always better!

 

[scoob8000]

Yep that is pretty common. Many that tried TPU went back. Faster isn't always better!

Man if that's the case I'm so disappointed for wasting money.

Some of these detections are beyond useless.

Car drives by, it's not detected but 20 or so people are, where there's nothing resembling a person.


Hopefully someone can chime in with a tweak or setting to try.

 

[wittaj]

The tweak is to remove the TPU and use the CPU LOL

What times were you getting with CPU? If the TPU isn't saving you more than 2 seconds, what is more important speed or accuracy?

 

[scoob8000]

The tweak is to remove the TPU and use the CPU LOL

What times were you getting with CPU? If the TPU isn't saving you more than 2 seconds, what is more important speed or accuracy?

Anywhere from 800-1500ms.

 

[wittaj]

I wouldn't think for most use cases 1.5 seconds is too long to wait for accurate detection?

 

[scoob8000]

I wouldn't think for most use cases 1.5 seconds is too long to wait for accurate detection?

I guess not. I got caught up in numbers.

Not like I need realtime AI for my usage. Most alerts I don't even look at until the next day. The only timely ones are stills sent via pushover while we're in away mode. But no real harm in that being delayed a few seconds.

I gave up on the coral and went back to CPU. The AI detection with is useless.

 

[MikeLud1]

I guess not. I got caught up in numbers.

Not like I need realtime AI for my usage. Most alerts I don't even look at until the next day. The only timely ones are stills sent via pushover while we're in away mode. But no real harm in that being delayed a few seconds.

I gave up on the coral and went back to CPU. The AI detection with is useless.

Try the Object Detection (YOLOv5 .NET) module, this will use whatever GPU you have.

 

[scoob8000]

Try the Object Detection (YOLOv5 .NET) module, this will use whatever GPU you have.

View attachment 207846

I'm giving that one a shot vs the 6.2 on CPU. Times seem to be about the same but I'll let it run for awhile.

I don't have a dedicated gpu, just whatever is built into the dell/i7 processor I've got. Running it shows DirectML GPU. Assuming that's correct?

Cheers!

 

[MikeLud1]

I'm giving that one a shot vs the 6.2 on CPU. Times seem to be about the same but I'll let it run for awhile.

I don't have a dedicated gpu, just whatever is built into the dell/i7 processor I've got. Running it shows DirectML GPU. Assuming that's correct?

Cheers!

Also make sure you have not checked Use main stream if available. Having it enabled does not improve accuracy it just slows down the detection.

 

[scoob8000]

Also make sure you have not checked Use main stream if available. Having it enabled does not improve accuracy it just slows down the detection.

View attachment 207853

Ahh. I do have use main stream checked. I don't know if that was the default or years ago I just thought it would be better.

I guess with the substream I can bump up my number of images like yours.

Currently I'm only doing 2 pre trigger, and 3 post trigger. Assuming more will get better accuracy?

[EDIT]
Performance seems to be best with .net, and CPU. Substream images are a bit faster to process. Around 400-500ms each vs 1200-1300ish for the main stream