2026 - Master Thesis - Algorithm Compression
ETH AI Center, UZH AI, PRS
- 2026
- 2023 - Seeing a Rose in Five Thousand Ways
- 2026 - VGGT-SLAM 2.0: Real-time Dense Feed-forward Scene Reconstruction
- 2026 - Reward-Conditioned Reinforcement Learning
- 📍 2017 - One-Shot Imitation Learning
Products
- Nvidia GEAR
- Matic Robots, home, level-5
- Rovex Technologies, hospital
- Fourier
- Flow, city navigation
- Taalas Inc.
Task Definition
The essence of LiDAR-free technology can be summarized as: transforming sparse measurements of the physical world into dense geometric inference.
- 📍 2020 - Convolutional Occupancy Networks
- 2022 - Scalable Diffusion Models with Transformers, William Peebles, DiT
References
- 2026 - High-Dimensional Probability
- 2025 - Visual Chronicles: Using Multimodal LLMs to Analyze Massive Collections of Images
- 📍 2026 - SteerVLA: Steering Vision-Language-Action Models in Long-Tail Driving Scenarios
Detection and Tracking Algorithms
| Layer | Technical Components | Application |
|---|---|---|
| Detection Backbone | RT-DETR | Transformer-based object detection producing spatially consistent bounding boxes and feature embeddings for downstream multi-view association. |
| Temporal Association | Query-Based Tracking (e.g., MOTR-style) | Maintains persistent object identities across frames using learnable spatiotemporal queries instead of heuristic matching. |
| Multi-View Correspondence | Epipolar Geometry | Filters cross-camera matches using the Fundamental Matrix to enforce geometric consistency before triangulation. |
| 3D Reconstruction | Triangulation + PnP | Recovers metric 3D positions from validated multi-view 2D detections and refines camera pose estimates. |
| Global Optimization | Bundle Adjustment | Minimizes reprojection error jointly over camera poses and object trajectories to achieve globally consistent 4D reconstruction. |
| Dynamic Motion Modeling | Motion Decomposition | Separates object motion from camera motion to stabilize optimization under dynamic scenes. |
| Spatiotemporal Refinement | Uncertainty-Aware Optimization | Weighs correspondences by confidence scores to improve robustness under occlusion, noise, and adverse weather conditions. |
| Identity Persistence | Cross-Camera Re-Identification | Uses learned feature embeddings to maintain consistent object identities across disjoint camera views. |