GeRaF 2.0 uses radar to reconstruct 3D objects hidden inside enclosures
Radar signals see through solid walls to create 3D shapes of hidden objects.
Reconstructing 3D geometry from radio frequency (RF) signals has been notoriously difficult due to the lensless nature of radar, which yields low spatial resolution and high noise. Unlike light, RF can penetrate occlusions, making it ideal for non-line-of-sight (NLoS) imaging. However, existing neural reconstruction methods for enclosed scenes often suffer from unstable optimization and ambiguous signed distance fields (SDF).
GeRaF 2.0, presented at CVPR 2026, solves this by explicitly modeling the LoS geometry outside the enclosure to guide RF propagation into the hidden region. This unified LoS+NLoS approach provides physical constraints that stabilize training and produce accurate zero-level sets. The result is a robust system capable of reconstructing fine 3D surfaces of objects hidden inside boxes, setting a new SOTA for radar-based geometry recovery.
- GeRaF 2.0 integrates visual LoS priors to guide radar propagation into NLoS regions for 3D reconstruction.
- Achieves stable training and physically consistent geometry, overcoming noisy surfaces and SDF ambiguity.
- Accepted at CVPR 2026, it sets a new state-of-the-art in RF-based 3D reconstruction of hidden scenes.
Why It Matters
Enables non-invasive 3D scanning of occluded objects for security, autonomous driving, and robotics applications.