This is what I'm looking for - I keep getting false positives on my weber under it's cover, or my son's black plastic motorbike etc. And the problem is if I raise the detection threshold high enough to avoid false positives then getting a good true positive becomes too hard. Plus both these things move around a lot in our back yard so trying to set filtered zones doesn't work. Feels like it's definitely worth trying frigate+ at least once.
How did you find the process of training the models?
At first, the process was pretty cumbersome, but I started a while ago and it’s improved a lot over time. Now, once you submit images to Frigate+, a separate model runs on them so that by the time you reach the training stage, most of what you’d want to select is already “pre-selected.” You mainly just tweak boundaries and maybe label one or two things. The biggest hassle is when training on images with lots of background items that get labeled, you have to label everything you’ve set as a possible label. For example, if you want to label cars only in a specific zone, you still have to label cars in all zones. I had trouble with a car parked on a neighbor’s driveway that I had to label every time, even though I wasn’t interested in it. The pre-labeling feature has made that much easier to deal with.
However, first things first, start with the base Frigate+ model, as that has already been trained on security camera footage, rather than the basic models which are trained on normal photos. So you might immediately see an improvement just by using the base model. Then you can try fine-tune using your own images.
Get used to the keyboard shortcuts, as it helps a LOT with speed. Over the last 2-3 years I have trained like 6000 images, but I've probably overdone it a little, to be honest. In the beginning, I would submit many similar images, which I now realise is just overfitting the model, and isn't necessary. So try and be selective with what you submit. There is some guidance in the documentation about what is a decent amount to submit between trainings.
But yeah - give it a shot. It will definitely help. Overall, the process has improved and is quite efficient. I'm mostly only submitting extremely edge cases in bad light and far corners of the property nowadays. And trying to train the AI that my crawling kid is not a dog...
