Feedback specificity

[yashishandilya]

Deciding what do we want to say, once we get distances between different points on te body

[sammypatel06]

Side note: There's two different standards for "good" posture; static and dynamic. Dynamic is posture when moving, static is posture when still. We are most likely only dealing with static posture, so all data should be recorded relative to a rating with a standard for a static posture in mind. (i.e. don't research dynamic posture). This does also mean we have to decide if we are rating a standing or sitting posture (or maybe have a toggle, or maybe detect automatically).

What is the "Data"

• MediaPipe gives us tracked body joints, so coordinate positions on a body
• Will also have angles between coordinates, distances, possibly curvature of parts such as the spine

Objective Feedback (Quantitative)

• Height
• Weight (guessing)
• Age (also guessing)
• Angles, distances

Subjective Feedback (Qualitative)

• The following is "good posture" and ratings/feedback should be made based on how well the user aligns to each characteristic
• Neutral spine alignment
  • Spine has 3 main curves: cervical (neck), thoracic (upper back), and lumbar (lower back)
  • Good posture maintains these curves and doesn't flatten or exaggerate them
  • Head should be balanced so ears line up roughly over shoulders
• Relaxed, level shoulders
  • Relaxed as in not hunched or pulled up toward ears
  • Open, rather than rounded forward
  • Level as in one isn't significantly higher than the other
• Aligned Hips and Pelvis
  • Not overly arched forward or tucked under
  • Standing: hips are level with spine, weight evenly distributed across both legs
  • Sitting: Thighs parallel to floor, both feet planted
• Balanced Weight Distribution
  • Standing
    • feet shoulder-width apart, knees soft (not locked), weight evenly distributed across feet
    • Chest open, chin parallel to floor
  • Sitting
    • Feet flat on floor (or footrest), knees at about 90 degrees
    • Lower back supported
• Minimal Muscle Tension
  • Not sure if we can "measure" this one
  • No excessive strain
  • Abs and back muscles aren't overworking to support, or being tensed aggressively
• Symmetry
  • Probably the easiest to start with
  • Not excessively twisting or leaning to one side, esp. in shoulders or hips
  • Not crossing legs (could probably just give an on-screen warning)

How much Data we need

• Simple
  • 1000+ frames labeled "good"
  • 1000+ frames labeled "bad"
  • ~30-50 people, multiple postures
• Moderate
  • Numerical score of posture
  • ~500 frames that are "good" or "bad" in each specific category
  • Multiple posture conditions, several more people
• High
  • 10s of thousands of frames with several levels of variability
  • Not realistic for our scale
• Too few samples => model doesn't generalize
• Too many samples => unbalanced data if samples aren't diverse
• Generally better to have lots of diverse data than not much of redundant data.

Specifying the Data

• Binary labeling (good vs. bad)
  • Easy to collect, less nuanced feedback
• Categorical (Good, slight slouch, severe slouch, etc)
  • More specific feedback, requires more data and more precision in training
  • Can easily be humanized into a small summary output
• Numerical Rating (1-10)
  • Tricky to label consistently, will require well-defined guidelines
  • Requires the most data

Outputting the Data

• Can output quantitative data so the user has numbers to attach to the output
• Can provide a small summary of large issues
• Can provide a small second paragraph of ways to fix the issues identified
• Can provide a rating of which issues are more severe and which ones are not
• Separate data by category (spine, shoulder, slouch, etc)
• Can record the user for about 5-10 seconds (or maybe even 30) to score posture over time; after the time period, the camera shuts off and the user is presented with a screen of feedback
• Another approach is giving the user a graph of "issues" that updates in real time as they adjust their posture