PolarDepth: Polarization-Guided Monocular Depth for Visual Odometry

Naitri Rajyaguru Tianfu Wang Aryan Tajne Botao He Jiayi Wu Cornelia Fermuller Christopher A. Metzler Yiannis Aloimonos
IEEE Robotics and Automation Letters (RA-L) — 2026
PolarDepth teaser

Glass surfaces remain challenging for indoor robot perception: depth sensors and standard RGB monocular frameworks consistently fail on transparent, low-texture, and reflective regions. To this end, we present PolarDepth, a polarization-enhanced monocular depth framework for glass-dominant environments. Using a single polarization sensor, we obtain a standard RGB image together with polarization cues, which we encode as a three-channel “Polar-RGB” input. Our network independently predicts depth from RGB and Polar-RGB and fuses them with a learned per-pixel reliability gate that trusts RGB in diffuse regions while emphasizing polarization cues on reflective glass. Designed to integrate with RGB-trained foundation depth models, PolarDepth substantially reduces tracking errors in glass-walled environments for downstream visual odometry.

@article{rajyaguru2026polardepth,
  title={PolarDepth: Polarization-Guided Monocular Depth for Visual Odometry},
  author={Rajyaguru, Naitri and Wang, Tianfu and Tajne, Aryan and He, Botao and Wu, Jiayi and Fermuller, Cornellia and Metzler, Christopher and Aloimonos, Yiannis},
  journal={IEEE Robotics and Automation Letters},
  year={2026},
  publisher={IEEE}
}