fitness_system_models

Model Card: PatchTST for Sports Motion Analysis

This model card provides detailed information about the PatchTST-based classification model used for detecting postural errors in exercises like Deadlift and Benchpress.

Model Details

• Model Name: PatchTST Classifier for Sports Form Analysis
• Model Type: Time Series Transformer (PatchTST)
• Developer: CatsLab (National Cheng Kung University)
• Date: March 2026
• Task: Multi-label classification of exercise technique and error detection.
• Supported Sports: Deadlift, Benchpress.

Architecture

The model is based on the PatchTST (Patch Time Series Transformer) architecture, which is specifically designed to handle longitudinal or sequential data by segmenting the time series into "patches" and processing them as tokens in a Transformer.

• Patch Length: 10
• Embedding Dimension: 256
• Number of Attention Heads: 4
• Number of Encoder Layers: 4
• Dropout Rate: 0.3
• Classification Head: LayerNorm followed by a Linear layer mapping to the number of classes.

Intended Use

• Primary Use: Automated analysis of resistance training form using pose estimation data (2D/3D coordinates or joint angles).
• Primary Users: Sports biomechanics researchers, fitness coaching app developers, and strength & conditioning professionals.
• Out-of-Scope Use: Medical diagnosis of injuries or real-time clinical gait analysis without proper validation.

Training Data

The model is trained on specialized sports datasets:

1. MOCVD (Deadlift): 1,109 sets including 8,506 repetitions from 156 subjects. Includes multi-view camera data and 3D reconstructed keypoints.
2. Benchpress Dataset: JSON-formatted trajectory data focusing on wrist and barbell movements.

Augmentation Techniques

To improve robustness, the following augmentations are applied during training:

• Time Stretching: Randomly stretching or compressing the sequence length by a factor of 0.8x to 1.2x.
• Gaussian Noise: Adding subtle noise ($\sigma=0.01$) to the input coordinates to simulate sensor or estimation variance.

Training Procedure

• Loss Function: BCEWithLogitsLoss (supports multi-label detection for overlapping errors).
• Optimizer: Adam ($LR = 10^{-4}$).
• Learning Rate Scheduler: Warmup Cosine Decay (5 warmup epochs, 100 total epochs).
• Input Length: 110 frames (Deadlift), 100 frames (Benchpress).
• Hardware: Trained on NVIDIA GPUs via CUDA.

Metrics & Evaluation

The model's performance is evaluated using:

• Macro F1-score: To account for class imbalance (crucial for error detection where 'Correct' samples often dominate).
• Accuracy: Overall prediction accuracy across all labels.
• Inference Time: Measured in seconds per sample.
• Multi-label Confusion Matrix: Providing class-specific False Positive and False Negative rates.

Factors & Bias

• Subject Variance: Performance may vary based on subject height, weight, and individual movement styles.
• Camera View: The MOCVD dataset is multi-view; however, performance might decrease if only an unoptimized or obstructed view is used.
• Pose Estimation Quality: The model is a "downstream" classifier; its performance is highly dependent on the accuracy of the upstream pose estimator (e.g., YOLOv11-Pose).

Ethical Considerations

• Privacy: Training data involves human subjects. All data used should be anonymized and handled in accordance with institutional privacy standards.
• Safety: AI-based feedback should not replace professional supervision for beginners performing high-load resistance training.

Caveats and Recommendations

• Normalization: Ensure that input sequences are normalized (Min-Max or Z-score) as expected by the model's preprocessing pipeline.
• Thresholding: The default classification threshold is 0.5. For critical error detection (where safety is paramount), practitioners may consider lowering the threshold to favor Recall over Precision.