Data Acquisition

- Download AIST++ & CREMA-D

- Extract keypoints with Mediapipe

Annotation / Clustering

- Annotate AIST++ subset (happy, sad, neutral) or Unsupervised clustering (k=3)

Feature Extraction

- V-JEPA v2 motion embeddings

- SMPL 3D pose features

- Essentia music features

Model Training

- Fine-tune V-JEPA v2 on AIST++ (self-supervised)

- Fine-tune on CREMA-D (emotion classification)

Emotion Classification

- Logistic regression on combined features

- Predict happy, sad, neutral

Evaluation

- Accuracy, F1-score on test set

- FID, BeatAlign scores

- User study feedback

AI Agent (Stretch Goal)

- Display emotion predictions

- Suggest movement improvements