- All old code from 2023-2024 is in old_code/
- Feb (or onwards) submission's work will be done in various root dir folders.
-
Feel that the strategies laid out in the original EPQ needs to be re-evaluated.
-
New strategies we can take up:
- Use some type of baseline (mediapipe, or mediapipe face mesh) -> use that to compute metrics like pitch etc.
- Feature extractor -> classification head
-
Fix current strategy with:
- Replacing things like “30 px” with T * face_height or delta/face_height.
- Adding persistence in some way, for example alert only if bad posture for k consecutive seconds/frames.
Then we can compute and compare metrics like accuracy, precision and recall, runtime (fps/cpu usage?), and how robust it is in terms of handling changes to lighting, distance or the camera angle.