Neural Geometry Reconstruction from Radio Frequency Signals
Event details
| Date | 21.07.2025 |
| Hour | 10:00 › 12:00 |
| Speaker | Jiachen Lu |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Alexandre Alahi
Thesis advisor: Prof. Haitham Al Hassanieh
Co-examiner: Prof. Mathieu Salzmann
Abstract
I will present a novel neural geometry reconstruction system that leverages radio frequency signals for close-range 3D reconstruction. Unlike RGB or LiDAR-based methods, RF sensing is robust to poor lighting and occlusion but suffers from low resolution and noise due to its lens-less imaging nature. While lenses in RGB imaging constrain sampling to one-dimensional rays, RF signals propagate through the entire three-dimensional space, introducing significant noise and leading to cubic complexity in volumetric rendering. Moreover, RF signals interact with surfaces via specular reflections, requiring fundamentally different modeling. To address these challenges, I introduce digital filter-based rendering to suppress irrelevant signals, implement a physics-based RF volumetric rendering pipeline, and propose a novel lens-less sampling and lens-less alpha blending strategy that makes full-space sampling feasible during training. By learning signed distance functions, reflectiveness, and signal power through multilayer perceptrons and trainable parameters, I successfully reconstruct high-resolution object geometry from millimeter-wave signals in real-world settings.
Selected papers
Exam president: Prof. Alexandre Alahi
Thesis advisor: Prof. Haitham Al Hassanieh
Co-examiner: Prof. Mathieu Salzmann
Abstract
I will present a novel neural geometry reconstruction system that leverages radio frequency signals for close-range 3D reconstruction. Unlike RGB or LiDAR-based methods, RF sensing is robust to poor lighting and occlusion but suffers from low resolution and noise due to its lens-less imaging nature. While lenses in RGB imaging constrain sampling to one-dimensional rays, RF signals propagate through the entire three-dimensional space, introducing significant noise and leading to cubic complexity in volumetric rendering. Moreover, RF signals interact with surfaces via specular reflections, requiring fundamentally different modeling. To address these challenges, I introduce digital filter-based rendering to suppress irrelevant signals, implement a physics-based RF volumetric rendering pipeline, and propose a novel lens-less sampling and lens-less alpha blending strategy that makes full-space sampling feasible during training. By learning signed distance functions, reflectiveness, and signal power through multilayer perceptrons and trainable parameters, I successfully reconstruct high-resolution object geometry from millimeter-wave signals in real-world settings.
Selected papers
- Wang, Peng, et al. "NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction." Advances in Neural Information Processing Systems 34 (2021): 27171-27183. (https://proceedings.neurips.cc/paper/2021/file/e41e164f7485ec4a28741a2d0ea41c74-Paper.pdf)
- Liu, Yuan, et al. "Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images." ACM Transactions on Graphics (ToG) 42.4 (2023): 1-22. (https://dl.acm.org/doi/pdf/10.1145/3592134)
- Borts, David, et al. "Radar fields: Frequency-space neural scene representations for fmcw radar." ACM SIGGRAPH 2024 Conference Papers. 2024. (https://dl.acm.org/doi/pdf/10.1145/3641519.3657510)
Practical information
- General public
- Free
Contact
- edic@epfl.ch