Shape from Polarization of Thermal Emission and Reflection

TL;DR

This work tackles shape-from-polarization for transparent materials by leveraging Long-Wave Infrared ($8-14\,\mu\mathrm{m}$) imaging, where most materials are emissive yet challenging due to reflection and sensor artifacts. It introduces a polarization model that jointly accounts for surface emission and environmental reflection, a calibrated LWIR polarimetric-imaging pipeline, and a physics-based learning method that trains a neural network on a synthetic dataset generated from the model. The authors implement a prototype and introduce ThermoPol, the first real-world LWIR SfP benchmark, demonstrating that learning-based normal estimation achieves superior accuracy (mean angular error around $10^{\circ}$ across materials, including transparent ones) and generalizes well beyond synthetic data. Key contributions include the emission+reflection polarization model, a robust Stokes-vector reconstruction pipeline, and a Transformer-augmented UNet that estimates per-pixel surface normals in view-space, enabling practical passive 3D reconstruction for challenging materials. The approach expands LWIR SfP applicability to translucent and opaque materials with spatially varying albedo and offers a path toward faster, more robust passive 3D sensing in realistic environments.

Abstract

Shape estimation for transparent objects is challenging due to their complex light transport. To circumvent these difficulties, we leverage the Shape from Polarization (SfP) technique in the Long-Wave Infrared (LWIR) spectrum, where most materials are opaque and emissive. While a few prior studies have explored LWIR SfP, these attempts suffered from significant errors due to inadequate polarimetric modeling, particularly the neglect of reflection. Addressing this gap, we formulated a polarization model that explicitly accounts for the combined effects of emission and reflection. Based on this model, we estimated surface normals using not only a direct model-based method but also a learning-based approach employing a neural network trained on a physically-grounded synthetic dataset. Furthermore, we modeled the LWIR polarimetric imaging process, accounting for inherent systematic errors to ensure accurate polarimetry. We implemented a prototype system and created ThermoPol, the first real-world benchmark dataset for LWIR SfP. Through comprehensive experiments, we demonstrated the high accuracy and broad applicability of our method across various materials, including those transparent in the visible spectrum.

Figures (15)

• Figure 1: Comparison of the polarization states in the visible and LWIR images. From top to bottom: radiance $L$, DoLP $\dolp$, and AoLP $\aolp$. The polarization state in the visible spectrum is significantly affected by transparency and surface albedos, compared to the LWIR spectrum.
• Figure 2: Radiance of a combined ray of the emitted component $T L_E$ and the reflected component $R L_R$.
• Figure 3: Reflectance and transmittance versus emergence angle $\theta$. Refractive index $\eta = 1.8$. For each of the $\mathit{p}$- and $\mathit{s}$-polarization components, the sum of reflectance and transmittance is always equal to 1.
• Figure 4: Our model of LWIR polarimetric imaging system.
• Figure 5: Characteristics of our imaging system. Left: Linear relationship between blackbody radiation differences and pixel intensity values across various temperature pairs ($\temp_\alpha$ = 23, $\temp_\beta$ = 20, 35, 50, 65, 80). This linearity validates our calibration approach based on differential measurements. Right: Polarization-dependent attenuation of the camera's microbolometer sensor, represented by gain factor $\mcampsi(\polangle)$ for rays with AoLP $\polangle$. The observed variation of approximately 5% confirms the need for polarimetric correction in our imaging pipeline.