A bug is when a particular feature is not working as it is specifically designed. This, unfortunately, is working exactly as they designed it. Is the design flawed? Yes. I believe so. And I believe it is because of the specific type of AI object anchoring they are using.
The method Wyze currently uses is a tradeoff between fine boundary constraints and speed of detection.
The current AI Wyze developed in house appears to be a single stage AI model that uses basic bounding boxes to anchor objects. Single stage models are much faster than two stage models, however they seriously lack in the ability to distinguish the fine boundaries of objects or multiple grouped objects. But, because of their speed, they are preferred for real time applications such as Wyze is producing.
The Wyze AI uses an incredible amount of real estate when drawing that massive bounding box rectangle over the object rather than using only the pixels within which the object is residing. This is the heart of the flawed design. The bounding box used to anchor the object is much too large and significantly decreases the effectiveness of using the DZ while increasing the effectiveness of the Overlap logic. This results in the an increase of true AI detections in areas designated for no detection. If they were to tighten up the bounding box anchoring, it could very well improve DZ accuracy but may decrease Overlap logic accuracy.
There are multiple examples of better models to anchor objects rather than using bounding boxes which introduce non-object area into the Included DZ. But most of these require far more computational power and also more time to process. Wyze is recording video at 15 and 20 fps. The AI model applying the object anchoring logic needs to be able to keep up with that frame rate. Many of the most accurate AI models using far more sophisticated object anchoring logic only run at between 5 and 10 fps and would be far too slow to provide real time AI Object Detection. They also require far more server horsepower.
The issue being discussed here is far more complicated than “fix it, it’s broke”. There are layers upon layers of dynamic limiting factors that must be considered when even the smallest modifications are made to the logic. I certainly want the AI model to be better. But thru understanding even the smallest bits about the technology, its capabilities, and it’s complexities, I can begin to understand why it is proving difficult for Wyze to dial in.
You have already read my response to that Wyze decision. I saw your like . Thank you.
That is what Direct Messaging is for. User to User direct. Not public. Most users in high traffic residential areas don’t have cams placed in areas that aren’t already viewed by the public on a daily basis or contain security sensitive information. All my security cams are highly visible to the public. I want it that way. Seeing the cams is a much greater deterrent than having an attack dog.
But, again, that is a personal decision. I can respect that.