Mistaking animals for people

[hspindel]

I am using CodeProject.AI v.295 in conjunction with Blue Iris running on a Win 11 machine with an old NVidia GPU (Quadro P620). The active model is YOLOv5 3.1 1.12.2. Custom models listed in Blue Iris include ipcam-animal, ipcam-combined, ipcam-dark, ipcam-general, and license-plate. Blue Iris is set to confirm alerts with AI and is set to alert only for person and animal. The model size is medium.

If this request is missing other important configuration information, please let me know.

CodeProject.AI very consistently reports that it is seeing people when animals wander through the frame. In a few months of observation, many animals (dogs, cats, raccoons, even flying birds) have come by and every single one of the was reported as a person. In contrast, vehicles are correctly recognized.

Is there something else I need to do to correctly distinguish animals from people?

Thank you.

[VideoDad]

You might try the YOLOv5 .NET module. It's a little more efficient with those custom ipcam models.

Also, are you just using all the custom models on all cameras? I generally like to pick a single custom model per camera. I therefore focus on using ipcam-general (person, vehicle) unless I want a little more granularity then I choose ipcam-combined. The dark objects are now in the other models so that's not necessary, IMO. And if you have a wildlife only camera, then use ipcam-animal.

You can turn on AI data logging for a camera (Camera Settings > AI tab) to help with diagnosing any issues. Once a clip is detected with an object, you can then hold down 'Ctrl' and double-click on the clip. That will bring up the AI inspector and you can see what custom models are being called, what objects they returned, the confidence levels, etc. That should help you determine if AI is doing what you expect.

That being said, I've always found animal detection to be spotty. I've seen birds and cats as people, and someone tying their shoe detected as an animal. So try to tune your AI first, but you might still some false positives for animals. AI isn't perfect.

[tbc725]

You can try to change the model size from medium to large. That might help the issue you are running into. Or use YOLOv8 and use larger models as needed.