#final03

We used MediaPipe to record 33 joints in each frame of every 2-second video.

1. Coordinate Normalization (Scale Invariance)

Raw motion data is often inconsistent due to varying camera distances and angles. To address this, I implemented a scale-invariant normalization logic:

The Logic: Using the body’s relative height (ref\_height\_px) as a reference, I transformed the 33 MediaPipe keypoints from pixel coordinates into a normalized, body-centric coordinate system.

The Formula:

Result: This ensures that movement trajectories remain mathematically consistent regardless of whether the dancer is close to the lens or further away.

2. Multimodal Data Integration (Standardized CSVs)

I have exported the processed data into structured CSV files for every 2-second clip. These files serve as a comprehensive "digital fingerprint" of the dance by aligning multiple data layers:

Motion Data: Precise X, Y, and Z coordinates for all 33 joints.

Audio Features: Synchronized audio metadata, including RMS (energy) and beat_strength.

3. Data Pruning and Quality Control

To ensure the integrity of the generative geometry in the next phase, I conducted a rigorous cleaning of the dataset:

Full-Body Filtering: I manually and algorithmically removed any clips where the dancer’s full body was not visible (e.g., feet or limbs moving out of frame).

Integrity: This step is crucial for maintaining continuous motion trajectories, preventing "glitches" when the data is later lofted into 3D forms in Grasshopper.

4. outputs：