CodeProject AI insane lag

[BCMguy]

Hello,

I just got CodeProject AI setup and am experiencing significant delays in processing the images. I'm talking like 1500+ ms. This seems to be way beyond what others are experiencing - any advice on things to check? CPU jumps sometimes to 100% while processing images too.

I7-4770 processor with 16GB ram
Latest BI and AI versions
9 camera feeds
Face rec off
Using IPCam custom models, cancel on banana, 15 real time images and 750ms each

Also, under BI Status, nothing appears under the AI tab - how do I get data to populate here?

Thank you for any advice!

 

[wittaj]

That is a 4th gen, so without a GPU the times will be slow, especially if you are using on more than one or two cams.

 

[BCMguy]

That is a 4th gen, so without a GPU the times will be slow, especially if you are using on more than one or two cams.

Good to know... I do see a GPU % flash up during triggers, that must just be the build in one I assume? Perhaps I'll need to look into a separate card. Lots of learning on my end -

 

[truglo]

15 real time images

As a data point, my intel 3470 with 8gb with otherwise similar settings using ipcam-combined averages 300-500msec, but I'm setup with 3+ images. If I used 15 images I would expect it to take 5x longer. In my case 3+ is plenty... the triggers are setup where whatever object is usually very easy to confirm with just 1 or 2 images to AI. If you really need 15 images for reliability, you may need better hardware IDK. Also, are you using all of the models, not just -combined? My heaviest cam is my LPR, which has combined, license-plate, and open ALPR... that one can take 500-800msec if there are plates involved. I could remove combined from my lpr to speed it up, but it has to serve multi role for now.

Also, under BI Status, nothing appears under the AI tab - how do I get data to populate here?

With that AI tab open, go back to the main window and open up one of your trigger/alert clips that was previously analyzed by AI (confirmed or cancelled, doesn't matter). Now go back to the AI tab, and the details for that clip will be shown.

 

[BCMguy]

Thank you for that info! I'll drop the images analyzed and edit the models being used. I believe I've been using combined, dark, and animal. In many cases it seemed like the default list was doing better, but I know it takes a lot of tweaking.

I'll report back!

 

[truglo]

Combined I think includes animal, so you can at least eliminate animal. Also I haven't used dark, since combined seems to work for me just as well regardless if it is day or night. So you might also be able to eliminate dark... or maybe setup profiles and scheduling to switch between dark and combined at dawn/dusk.

 

[BCMguy]

Alright making those changes down to 4 images and dropping all but the Ipcam general list I'm down to about 400ms! I'm getting a lot more Nothing Found drops, but I'll try one of the other lists and see how they work out.

Still using main stream images. Any thoughts on the Apply to motion triggers only option?

 

[MikeLud1]

Still using main stream images

Try unchecking main stream, you will still get the same accuracy with faster times, it may drop your times to about 200ms

 

[truglo]

Any thoughts on the Apply to motion triggers only option?

If that is checked AI won't process any manual triggers. I leave it unchecked so I can use manual triggers for testing.

 

[BCMguy]

If that is checked AI won't process any manual triggers. I leave it unchecked so I can use manual triggers for testing.

Thank you. So far my times are around 2-300ms. Dramatic improvement!