ML & Computer Vision Disclosure

ML-Powered Features

KineticForm uses on-device machine learning and computer vision to enhance your fitness experience. This disclosure explains how these technologies are used and their limitations.

How We Use ML

Form Analysis

Our computer vision model analyzes your body position and movement during exercises using your device's camera. The system:

Uses Standard Analysis with Apple's Vision framework (18 joint points)
Uses Advanced Analysis with ARKit Body Tracking (91 joint points on supported A14+ devices)
Processes video locally on your device
Compares your form against general exercise guidelines
Provides real-time feedback and suggestions

Analysis is available across 27+ exercises including compound lifts, isolation movements, and bodyweight exercises.

Progressive Overload Suggestions

Our ML model recommends weight increases based on:

Your workout history and performance
Your selected progression algorithm
Rest times and recovery patterns
General strength training principles

ML Limitations

ML-generated feedback is not a substitute for professional instruction. The ML system has inherent limitations and may not always provide accurate or appropriate advice.

Our ML model may:

Miss form errors that are subtle or outside camera view
Provide generic feedback that doesn't account for your specific body type
Struggle in poor lighting or with obstructed camera views
Not account for pre-existing injuries or conditions
Occasionally provide incorrect or suboptimal suggestions

Always prioritize how your body feels over the app's suggestions.

Estimated Accuracy

The following are estimated accuracy ranges based on internal testing in controlled environments. Formal validation with certified trainers is ongoing. Real-world accuracy varies based on lighting, camera angle, body type, device placement, and clothing.

Exercise	Estimated Accuracy	Detail
Squat	~91%	Depth detection
Bench Press	~87%	Bar path tracking
Deadlift	~85%	Back posture
Overall	~88%	Estimated agreement rate

These figures are preliminary estimates and will be updated as we complete formal validation studies. Individual results may differ significantly.

Scientific References

Our feedback rules and training recommendations are informed by established sports science and exercise physiology research, including:

Foundational

Schoenfeld (2010) - Muscle hypertrophy and resistance training mechanisms (DOI: 10.1519/JSC.0b013e3181e840f3)
Kerksick et al. (2017) - Nutrient timing position stand (DOI: 10.1186/s12970-017-0189-4)
Schoenfeld & Aragon (2018) - Protein distribution in muscle-building contexts (DOI: 10.1186/s12970-018-0215-1)
ACSM (2009) - Resistance training progression models (DOI: 10.1249/MSS.0b013e3181915670)
Borg (1982) - Perceived exertion foundations (DOI: 10.1249/00005768-198205000-00012)
Kellmann (2010) - Stress/recovery monitoring and overtraining prevention (DOI: 10.1111/j.1600-0838.2010.01192.x)

Recent Evidence (2019–2025)

Refalo et al. (2023) - Proximity-to-failure and skeletal muscle hypertrophy meta-analysis (DOI: 10.1007/s40279-022-01784-y)
Schoenfeld et al. (2021) - Repetition continuum re-examination for strength and hypertrophy (DOI: 10.3390/sports9020032)
Schoenfeld & Grgic (2020) - Range of motion effects on muscle development (DOI: 10.1177/2050312120901559)
Schoenfeld et al. (2019) - Training volume dose-response for hypertrophy (DOI: 10.1249/MSS.0000000000001764)
Weakley et al. (2021) - Velocity-based training theory and application (DOI: 10.1519/SSC.0000000000000560)
Helms et al. (2016) - RIR-based RPE scale for resistance training (DOI: 10.1519/SSC.0000000000000218)
Vitale et al. (2019) - Sleep hygiene for optimizing recovery in athletes (DOI: 10.1055/a-0905-3103)

See our full citation list on the Scientific References page.

How It Works

Angle Calculation

We measure joint angles using 3-point geometry (e.g., shoulder-elbow-wrist for elbow bend). Angles are smoothed over 5 frames to reduce jitter.

Posture Rules

Each exercise has specific rules (e.g., "back angle < 45° in squat" to detect forward lean). Rules are based on established biomechanics literature. Calibration adjusts thresholds to your personal range of motion.

Rep Detection

We track movement phases (setup, eccentric, bottom, concentric, lockout) to count reps. A minimum 1-second pause is required between reps to prevent double-counting.

Fairness and Bias

We are committed to building ML models that work for everyone. We:

Test our models across diverse body types and fitness levels
Continuously work to reduce algorithmic bias
Welcome feedback if you experience inaccurate results
Update our models based on user feedback and research

If you notice the model seems less accurate for you, please let us know at support@kineticform.app.

Your Data and ML

ML processing happens on your device, not cloud servers

Your workout data is not used to train ML models

We do not share your data with third-party ML providers

You can disable ML features at any time in Settings

Responsible Use

For the best and safest experience:

Always warm up before exercise
Trust how your body feels over app feedback
Stop if you experience pain or discomfort
Consult a trainer for personalized instruction
Report any concerning model behavior through the app