Closed angad305 closed 6 months ago
Ah, so I didn't dive into how the cascading rule-logic works in the video. The system has a few levels of rules which are more or less confident and accurate at detecting a sleeping baby. If a more accurate rule is "confident", the system ignores the less accurate rules (e.g. the custom AI model that is trained when you click "Yes" or "No"). In this case, a more accurate detector is running, and the system isn't invoking the custom model.
So what's happening is: because it's detecting anatomical features: it's disabling the buttons. I can add to my TODO list to explore enable training even if the custom model is not running.
What's more concerning from your pic is that the overlayed circles where it thinks the face is, seems off. Does it always say, "Anatomical features detected, ignoring custom presence detection AI", even when your baby is out of frame? I may need to increase the confidence threshold of that "rule", so it is less likely to misfire.
Also, I'm not seeing an arrow in the "sleepometer", pointing to the system's confidence that your baby is sleeping. Do you ever see the arrow?
e.g.
Ah, so I didn't dive into how the cascading rule-logic works in the video. The system has a few levels of rules which are more or less confident and accurate at detecting a sleeping baby. If a more accurate rule is "confident", the system ignores the less accurate rules (e.g. the custom AI model that is trained when you click "Yes" or "No"). In this case, a more accurate detector is running, and the system isn't invoking the custom model.
So what's happening is: because it's detecting anatomical features: it's disabling the buttons. I can add to my TODO list to explore enable training even if the custom model is not running.
What's more concerning from your pic is that the overlayed circles where it thinks the face is, seems off. Does it always say, "Anatomical features detected, ignoring custom presence detection AI", even when your baby is out of frame? I may need to increase the confidence threshold of that "rule", so it is less likely to misfire.
i do get Understood the logic part. thank you for the explanation.
See below screenshot, the "Anatomical features detected, ignoring custom presence detection AI" is always there mostly.
Also, I'm not seeing an arrow in the "sleepometer", pointing to the system's confidence that your baby is sleeping. Do you ever see the arrow?
e.g.
the issue is on the big screen monitor when i use chrome, on mobile phone the arrow pointer is working fine.
Got it on the mobile vs desktop.. I didn't spend time making the web app portion responsive. Primary use case is mobile.
Regarding always seeing, "Anatomical features detected...", can you try going into main.py
and increasing the two confidence arg values?
For my crib environment, I found a lower value (looks like 0.3) worked well for me and my baby. But you may need to increase this value so that you aren't getting false-positives, if that makes sense. If it's triggering the MediaPipe pose-detection model, then it won't default to the new final catch-all model.
This reminds me, I should expose an easier way within the web app to enable users to more easily fine-tune these thresholds
Both the buttons for AI Confidence are disabled? Please see attached pic.
Is there some logic which is preventing them.