Traditional 2D input devices are not natural for interacting with 3D environments. This project replaces mouse-based control with temporal hand-motion understanding and validates whether manipulated designs remain physically plausible.
- Computer vision for robust hand landmark extraction.
- Deep learning for temporal intent recognition.
- Human-computer interaction for natural 3D manipulation.
- Lightweight design validation inspired by physics-aware optimization research.
Build an interactive system where users can create, move, rotate, scale, join, and delete 3D objects through hand signs, then receive instant structural feedback.
- Hand tracking: MediaPipe Hands.
- Temporal model: LSTM or Transformer.
- Rendering and interaction: Three.js.
- API and inference layer: FastAPI.
- Numeric processing: NumPy.
- Webcam stream capture.
- Hand landmark extraction (21 points per frame).
- Sequence buffering over a temporal window.
- Gesture intent classification.
- 3D action mapping.
- Physics heuristic validation and feedback.
- Week 1-2:
- Implement camera capture and landmarks.
- Display stable tracking.
- Week 3-4:
- Build dataset and train baseline LSTM.
- Evaluate model with confusion matrix.
- Week 5:
- Integrate model output with Three.js actions.
- Demonstrate create, move, rotate, scale, delete.
- Week 6:
- Add physics scoring and warning overlay.
- Validate behavior on predefined test scenes.
- Week 7:
- UX polish, latency reduction, final presentation and report.
- Single hand for first version.
- Primitive shapes only.
- Heuristic physics in place of full finite element simulation.
- Show real-time hand tracking.
- Create an object using a gesture.
- Grab and move object in 3D scene.
- Rotate and scale object.
- Trigger physics check and show stability/stress metrics.
- Demonstrate warning and suggestion output.
- Gesture intent recognition accuracy above baseline threshold.
- End-to-end interaction latency acceptable for real-time use.
- Physics feedback correctly flags intentionally unstable examples.