Project Roadmap
1 Project Roadmap
1.1 Seminar Phase (6 ECTS, ~180 hours)
1.1.1 Phase 1: Foundation & Familiarization
Goal:
- Get familiar with the NBV problem domain and read relevant literature
- Understand the Aria ecosystem — ASE data, SceneScript, Project Aria Tools, ATEK, EFM3D/ EVL
1.1.2 Phase 2: Oracle RRI
Goal: Implement an “oracle” Relative Reconstruction Improvement (RRI) computation method based on the GT meshes and semi-dense point clouds or compute dense PC from GT depth maps (problem: they are only covering actual trajectory)
1.1.3 Phase 3: RRI Prediction Network w/ EVL or Scene-Script backbone
Goal:
- Use pre-trained SceneScript or EFM3D/ EVL model to predict RRI for candidate views based on previous observations (partial point cloud, previous poses and calibrated video frames).
- Define training procedures and loss functions.
1.1.3.1 Architecture
- Encoder: SceneScript encoder or EVL backbone (frozen or fine-tuned)
- Candidate Encoder: View pose + frustum features, projection of partial PC into candidate view -> learnable encoder
- Predictor: MLP regressing RRI score
1.2 Master Thesis Phase (30 ECTS, ~900 hours)
1.2.1 Phase 4: Entity-Aware RRI
1.2.2 Phase 5: View Synthesis Integration
1.2.3 Phase 6: Learning-Based NBV Prediction
1.2.4 Phase 7: Fine-Tune EFM for NBV on Target Platform
- Project Aria Glasses: High-fidelity egocentric data
- Meta Quest 3: Inside-out tracking, depth sensor
- iPhone LiDAR: Portable, high-quality depth
1.2.5 Phase 8: Human-in-the-Loop System
Goal: Develop interactive AR interface for real-time NBV guidance
1.2.5.1 Components
- Entity Selection UI
- Tap entities in AR view
- Set importance weights
- Real-Time NBV Computation
- Streaming point cloud input
- Incremental updates of the scene representation
- Fast RRI and pose prediction
- View Guidance Overlay
- AR arrows/markers showing optimal viewpoints
- Distance and quality indicators
- Real-time voice2voice feedback
1.2.6 Phase 9: Real-World Deployment & Evaluation (Months 7-8, ~200 hours)
Goal: Deploy on mobile devices and validate in real environments