ipcams2 Computer Vision Project

The project is for automated processing of home video camera feeds. This dataset includes both daytime and nighttime (IR) images, typically from perspective of a typical camera.

I suggest splitting the dataset and training two models: one for daytime and the other for nighttime. The nighttime pictures have a single channenl while the daytime ones have three channels, this results in significantly different features being trained. I identify if the image has one or three channels using the following shell command: identify -colorspace HSL -verbose "$f" | egrep -q "(Channel 0: 1-bit|red: 1-bit)"

The images are full size, so different sized models can be created. I've been training at 608x608. It includes many null images which have in the past triggered a false positives.

The classes are simply the things of interest I've seen from my house. In general this is more useful than the standard yolo classes, such as Zebra. However, you may want to have bear or some other wildlife. I've found squirrels are too small for my cameras to reliable pickup and detect. The perspective and framing of content is quite different from typical stock photos, so I think it makes a lot of sense to train the model using only images from ipcams.

Ideally, I will make models available for the many different tools people are using for AI already, including: Deepstack / BlueIris MotionEye Frigate