A depth camera can mistake the scene behind a glass for the glass itself. The missing measurement is the transparent surface.[1]
Glass presents a peculiar opportunity. Light reflected from its surface can become partly polarized: its electric field favours particular directions of oscillation. A standard colour image mainly records brightness and colour. A polarization camera can also measure the orientation and strength of that directional preference. What often appears as annoying glare can therefore carry information.[2]
The camera does not need to rotate a filter for every measurement. A sensor can place four differently oriented polarizers over neighbouring pixels, at 0, 45, 90 and 135 degrees. Comparing their intensities reveals more than those four angles: together they allow the direction and degree of linear polarization to be estimated.[2]
From a reflection to a surface
Imagine tiny arrows sticking straight out of a surface. Those arrows are its normals: they describe which way each patch faces. If neighbouring arrows turn smoothly, the surface curves smoothly. If their directions change sharply, there may be an edge or a dent. Polarization measurements can help estimate these orientations; integrating their implied slopes reconstructs shape. LUCID’s documented shape-from-polarization demonstration uses this principle to inspect dents in a metal can.[3]
PolarDepth, presented by Wen Dong and colleagues at ICML 2026, starts with a frozen Depth Anything V2 model. It reconstructs shape from learned surface orientations and polarization physics, then anchors its scale to that initial depth estimate. Only transparent regions receive the correction.[1]
This division of labour is useful. A local slope says how a surface tilts; it does not by itself say how far the whole object stands from the camera. Shape and distance need compatible constraints. The physical measurement contributes information that an image’s visible pattern may never contain.[3]
On the test portion of its 2,878-scene PTOD dataset, transparent-region relative depth error averaged 0.142, against MODEST’s 0.252. This measure divides each absolute distance error by the reference depth before averaging.[1]
Measuring a surface the sensor cannot see
PTOD’s reference depths use aligned opaque replicas, with sensor measurements used to align and scale dense Depth Anything V2 predictions. Transfer remains difficult: after PTOD training, relative error rose to 0.651 on a HAMMER subset; training on both datasets reduced it to 0.179.[1]
The result points to an appealing direction for machine vision: measure a property of the world that reveals the missing structure, then give learning a physically meaningful way to use it. For glass, transparency does not mean an absence of evidence. Some of the evidence is carried in the light’s orientation.
Sources & further reading
- PolarDepth: Monocular Transparent Object Depth from Polar-Physics Priors Proceedings of Machine Learning Research · Accessed 18 Oct 2026
- Sony’s Polarized Sensor: Beyond Conventional Imaging LUCID Vision Labs · Accessed 18 Oct 2026
- 3D Depth from Surface Normals using a Polarization Camera LUCID Vision Labs · Accessed 18 Oct 2026