The images it uses for comparison, as explained in the link, are new ones, taken from the start of the recording process, after the PIR sensor triggers on detecting a moving heat signature. This snippet of video, probably two frames, a couple of seconds apart, is analysed, pixel by pixel, and compared to the mask. If the motion is outside of the masked area, then the recording will continue and be uploaded. If the motion is entirely within the masked area, the recording will cease, and what has been recorded will be discarded.
In theory this should cut down on false alerts, and probably does, but for sure it isn’t perfect, at least in my experience, and if comments from other community members are anything to go by.
If you post a video link of what you consider to be a false trigger, and a screen shot of your mask, maybe we can make out why you are getting unexpected triggers.