New CodeProject.AI Object Detection (YOLO11 .NET) Module

Mike (@MikeLud1),

I'm still on 5.9.9.89 running CPAI. I know your focus is on the built-in AI in Blue Iris v6. Is there a way to get your models for BI6 and use them in BI5 using CPAI and this YOLO11 .NET module?

In particular, what models have you regenerated for BI6 and where are they located? I couldn't find them in your GitHub.
 
VideoDad said:
Mike (@MikeLud1),

I'm still on 5.9.9.89 running CPAI. I know your focus is on the built-in AI in Blue Iris v6. Is there a way to get your models for BI6 and use them in BI5 using CPAI and this YOLO11 .NET module?

In particular, what models have you regenerated for BI6 and where are they located? I couldn't find them in your GitHub.
Click to expand...
You can download the models that come with Blue Iris from the below link. Change the file extension from exe to zip and then you can open it with WinRAR or WinZip. Most of the models are still my old models.

What model are you looking for. I just recently retired and want to start training newer models.


1788072904124.png
 
The one that I'm really looking for is IPcam-combined.onnx with the different vehicle types (truck, car, van, bicycle, etc.) and animals (cat, dog, bird, etc.)
 
MikeLud1 said:
They work with my ObjectDetectionYOLO11 module, I have not trained these models with YOLO11
Click to expand...
Right, I understand that. But you asked which one I'm hoping you'll retain for YOLO11 and that would be it. And I've seen prior requests for the same.

Anyway, I'll switch to the v5 version in the meantime.
 
Hello. I recently found this thread by chance looking for performance info for BI built in AI. I have a RTX 3050 6GiB card. I gave it a go, and using the v5.onnx ipcam-combined(day profile) / ipcam-dark(night profile), I am getting excellent inference times with the model size set to large and sending main stream images to CPAI. Now with BI having Tensor support, I tried switching back. I found that I am getting way slower round trip times in the BI log (90-200ms with yolo11.net/cpai and 400-700ms with blue iris built in AI --not using blue iris built in primary model at all). I am using the same models and settings for both cases. (max analyzers = 2, preload all models turned on with blue iris built in AI). Average system cpu useage also seems up by 5-10 percent on average when using BI built in AI. (I tried remvoing all but the two v5 models from the BI AI models folder incase preloading all the others would effect this). This doesnt sound right to me. I would think that with BI built in AI I would get faster round trip times seeing that the REST api is taken out of the picture) Any paths or things I should be looking at? Thanks all in advance for any suggestions.

Edit: I'm on BI 6.1.3.4. CPAI is 2.9.5. Win 11 all updates installed. i7-9700k with 30gb ram.
 
MikeLud1 said:
Do not use main stream images for AI, it does not improve accuracy, it only slows down the detection
Click to expand...
I will uncheck all the mainstream boxes on the cam settings in BI. I had tested back and forth in the past several times and it must not have effected inference times enough for me to bother. Reducing the times is a good thing, but I don't think that would account for the discrepancy in performance between BI and CPAI in my case. Your yolo11.net with ipcam-combined-v5 and ipcam-dark-v5 are still giving me better performance across the board. With BI adding WinML, my hope in moving away from CPAI was to remove one link in the chain and ease management of the system. CPAI installs fine today, but who knows in the future if that will be the case. Good thing for full system backups. Thanks!
 
MikeLud1 said:
Do not use main stream images for AI, it does not improve accuracy, it only slows down the detection.
Click to expand...
I have another question related to this. I have mainstream images turned on to get high quality images sent in alerts. Is this fixed now where I can use the substream for AI detection, but still get the HQ image for sending with the alert?
 
VideoDad said:
I have another question related to this. I have mainstream images turned on to get high quality images sent in alerts. Is this fixed now where I can use the substream for AI detection, but still get the HQ image for sending with the alert?
Click to expand...
I just switched a bit ago, and in UI3 when I right click on an alert and hit properties, the image size recorded is the smaller substream resolution. You scrached an old memory and one of the reasons in the past I had left mainstream images on was so that saved image would be at the full 8mp size. It's an unfortunate tradeoff still I think.
 
  • Like
Reactions: MikeLud1
You must log in or register to reply here.
Top Bottom