✅ Day 5 – From Paper to Practice: My First Experiment

After reviewing the HDVR (Hierarchical Dance Video Recognition) paper in depth,
I wanted to start building a simplified version of the pipeline using real pose data and my own code.

The original paper relies on 3D lifting and motion segmentation, but I decided to begin with the 2D keypoint side: extracting features like joint distances, angles, and velocities from pose sequences and visualizing motion.

🛠️ What I Built – 2D_Pose_Feature_Builder.ipynb

🎯 Purpose

• Reproduce the lower half of the HDVR pipeline, focused on body part movement from pose sequences.
• Lay groundwork for temporal modeling (LSTM/TCN) using only pose-derived features.

📂 Functionality

• Input: .json or .csv file of 2D pose keypoints (from MediaPipe or OpenPose)
• For each frame:
  • Compute joint-to-joint distances (e.g., wrist to elbow)
  • Calculate joint angles (e.g., elbow angle from shoulder–elbow–wrist)
  • Optionally compute velocity over time (joint displacement)
• Visualize:
  • Skeleton overlay by frame
  • Feature value sequences (time-series)

💡 Why This Matters

This experiment helped me:

• Understand which joints are stable vs. noisy
• See that even simple handcrafted features (distances, angles) encode a lot of motion semantics
• Identify where future smoothing or filtering would help
• Confirm that pose-only pipelines are viable for lightweight modeling

It was also important to validate that a full 3D lifting module isn’t strictly necessary for building a useful system, especially in constrained or real-time environments.

📊 My Model Setup (So Far)

No learning model yet – this was a feature engineering stage only.
But these outputs will soon be fed into a temporal model like:

• LSTM / GRU for frame-level genre classification
• TCN for learning local body part movement patterns
• Possibly Transformer-based encoder for attention over joint importance

🔭 Next Steps (aka Day 6 Plan)

• Integrate pose segmentation: split long videos into movement segments
• Build a sequence classifier for genre (e.g., hip-hop vs. waacking)
• Try movement encoding from multiple dancers simultaneously
• Optionally explore 3D lifting with mock or learned constraints

📝 Reflection

Implementing this part manually gave me a better grasp of why the HDVR paper’s hierarchical structure makes sense.
Rather than depending on raw pixel data or supervised 3D annotation, I can now construct explainable, modular pipelines based entirely on pose motion.

This also sets the foundation for a real-time dance feedback system, or downstream applications in fitness, rehab, or choreography assistance.