📚 https://doi.org/10.1038/s41598-024-83475-4

✅ Day 1 – Explainable Skeleton-Based Evaluation: From Reading to Application

Today I reviewed a paper that aligns perfectly with my long-term goal:
building interpretable movement evaluation systems based on pose estimation.
Unlike classification-only pipelines, this study dives into scoring, alignment, and explanation — all with a practical and reproducible approach.

🧠 What I Read – Explainable Skeletal Assessment (Scientific Reports, 2025)

🎯 Purpose

• Develop a motion evaluation framework that is:
  • Quantitative (scores performance),
  • Aligned (accounts for variation between performers),
  • Explainable (highlights which joints/motions influenced results).

📂 Key Components

1. Input: Skeleton sequences from both reference (expert) and target (trainee) movements
2. Alignment:
   • Spatial: Procrustes Analysis
   • Temporal: Dynamic Time Warping (DTW)
3. Modeling:
   • Decision Tree, Logistic Regression, LSTM used to predict skill level or score
4. Explanation:
   • SHAP values visualize which joints contributed most to final predictions

💡 Why This Matters

This paper helped me:

• Understand how motion scoring can go beyond classification into interpretable evaluation
• Realize the value of alignment methods in making comparisons fairer
• See how combining traditional ML models and SHAP can yield intuitive, robust outputs
• Think about how to make feedback actionable, not just accurate

The transition from black-box genre recognition to explainable scoring is essential for real-world use in sports, fitness, or dance.

📊 My Implementation Plan (So Far)

This paper will influence my next prototype module. Planned steps:

• Use 2D pose data (MediaPipe) instead of 3D
• Apply Procrustes + DTW to align a reference pose and a trainee video
• Score similarity via cosine + statistical features
• Later add model-based scoring (e.g., shallow tree or regression model)
• Integrate SHAP or heatmap-based visualization to show which joints “failed”

🔭 Next Steps (aka Day 2 Plan)

• Collect 2D pose data with more consistent structure
• Reimplement Procrustes + DTW alignment in my notebook
• Compare frame-level vs. sequence-level evaluation
• Try multi-sample evaluation: one reference vs. many attempts
• Visualize joint influence maps from SHAP or cosine error stats

📝 Reflection

This paper gives structure to my vague ideas about explainable dance evaluation.
It doesn’t try to “solve everything” with deep learning but instead shows how simple techniques + smart alignment + explainability can work together.
A big takeaway: use alignment not just for preprocessing but as part of the model logic.
Next, I’ll build a mini-pipeline using DTW-aligned 2D pose features and map frame-level scores back to visuals.

This could evolve into a feedback system for dance, fitness coaching, or rehab scenarios — just like I originally envisioned.