I've been using Blue Iris for a couple of years now. Running on a Core i7-9700k, Windows 11 PC with 16GB Ram. Blue Iris is installed on a SSD. Only running 5 cameras currently.

I think I have finally found a decent setup and have my system running well and has been very reliable. I want to share this information, but I am also looking for pointers on how to get the most out of Blue Iris. Some recent changes I have made to my setup...

1. Moved all my cameras to a separate VLAN.
2. I'm using a dual NIC setup on my Blue Iris Server. One interface as internet access, and the other is for my cameras that are on their own VLAN.
3. I installed a Google Coral TPU using a PCIe express slot.
4. I'm running CodeProject AI and using the Coral TPU for object detection.

Are there any Blue Iris experts out there that can recommend any further improvements? I know some of the settings will be specific to the installation and how things are running, but is there anything that looks incorrect or I could improve on.

[attachment=0]


Looks like I can only attach 3 pictures per post. Will add more next if possible.

I'm quite impressed how low the timing is to analyse; my object detections are anywhere in between 100-200ms using the medium model. My setup is a virtual machine with 4x CPU cores Ryzen 5800U with the TPU using CPI-E passthrough in hyper-v, so in theory it should be pretty good and low overhead.

I'm curious how can you run a custom model for object detection, when on your screen shots there is none displayed on the general AI configuration in BI. so far I know, the IPcam-combined is only available for Yolo. Could you please share a bit more about that?

nothinger wrote: ↑Sat Dec 30, 2023 3:52 pm I'm quite impressed how low the timing is to analyse; my object detections are anywhere in between 100-200ms using the medium model. My setup is a virtual machine with 4x CPU cores Ryzen 5800U with the TPU using CPI-E passthrough in hyper-v, so in theory it should be pretty good and low overhead.

I'm curious how can you run a custom model for object detection, when on your screen shots there is none displayed on the general AI configuration in BI. so far I know, the IPcam-combined is only available for Yolo. Could you please share a bit more about that?

I have been told that the custom models while using Coral TPU doesn't work yet. So while I may have custom model checked, I think it is using the default model.

nothinger wrote: ↑Sat Dec 30, 2023 3:52 pm I'm quite impressed how low the timing is to analyse; my object detections are anywhere in between 100-200ms using the medium model. My setup is a virtual machine with 4x CPU cores Ryzen 5800U with the TPU using CPI-E passthrough in hyper-v, so in theory it should be pretty good and low overhead.

I'm curious how can you run a custom model for object detection, when on your screen shots there is none displayed on the general AI configuration in BI. so far I know, the IPcam-combined is only available for Yolo. Could you please share a bit more about that?

I should also mention I am using the TPU in a PCI-e socket. It isn't the USB version. Not sure if that has any impact or not. I assume the extra delay you are seeing must be somehow related to the VM and passthrough? Still 100-200ms isn't bad.

douga wrote: ↑Fri Jan 05, 2024 8:46 pm

nothinger wrote: ↑Sat Dec 30, 2023 3:52 pm I'm quite impressed how low the timing is to analyse; my object detections are anywhere in between 100-200ms using the medium model. My setup is a virtual machine with 4x CPU cores Ryzen 5800U with the TPU using CPI-E passthrough in hyper-v, so in theory it should be pretty good and low overhead.

I'm curious how can you run a custom model for object detection, when on your screen shots there is none displayed on the general AI configuration in BI. so far I know, the IPcam-combined is only available for Yolo. Could you please share a bit more about that?

I have been told that the custom models while using Coral TPU doesn't work yet. So while I may have custom model checked, I think it is using the default model.

You would need to have the custom files in your folder to make any difference from what I understand as I tested this with the other models like Yolo and under custom they were showing. I read some messages from the chap that developed these models on github and he said he was working on them, but they were not available yet.

douga wrote: ↑Fri Jan 05, 2024 8:49 pm

nothinger wrote: ↑Sat Dec 30, 2023 3:52 pm I'm quite impressed how low the timing is to analyse; my object detections are anywhere in between 100-200ms using the medium model. My setup is a virtual machine with 4x CPU cores Ryzen 5800U with the TPU using CPI-E passthrough in hyper-v, so in theory it should be pretty good and low overhead.

I'm curious how can you run a custom model for object detection, when on your screen shots there is none displayed on the general AI configuration in BI. so far I know, the IPcam-combined is only available for Yolo. Could you please share a bit more about that?

I should also mention I am using the TPU in a PCI-e socket. It isn't the USB version. Not sure if that has any impact or not. I assume the extra delay you are seeing must be somehow related to the VM and passthrough? Still 100-200ms isn't bad.

I don't think it should be adding so much latency, I think your config must be different somewhere because the size of the model changes the latency a bit. If I use a smaller model it lowers the latency a bit, but not much. Also, I'm surprised you analyse 10 frames, and still get 30ms.