Haar Cascade accuracy/consistency degraded in firmware 4.3.1

With the latest firmware, the Haar cascade algorithm seems to be very inconsistent and less accurate than previously. With the image mostly stationary, the algorithm will cycle between detecting a face and not detecting one. This is a new behavior. Any thoughts?

It might be different lighting, the Haar cascades are very sensitive to lighting, that code hasn’t been changed for years. If you need face-detection, there’s a built-in tensorflow object detector (EI FOMO model) trained on faces, in the firmware now. It’s not released yet, so you’ll have to download and flash the latest development firmware from github (and use the example from github too).