Polarimetric Surface Normal Estimation for Enhanced Vehicle Pose Classification

Date of Award

8-15-2026

Degree Name

M.S. in Electrical and Computer Engineering

Department

Department of Electrical and Computer Engineering

Advisor/Chair

Brad Ratliff

Abstract

This paper investigates the use of polarimetric imaging for the extraction of surface normals to enhance object detection. While traditional intensity-based imaging often struggles with lighting direction, shading, and shadows, polarimetric data provide additional dimensions to recover complex surface geometries. The experimental setup in the Automated Remote Sensing Solar Simulation Lab at The University of Dayton uses a custom Blackfly Visible Multispectral Division of Time (DoT) polarimeter. Four different polarimetric orientations (0◦ , 45◦ , 90◦ , and 135◦ ) across three spectral RGB channels are captured. The data were collected from various colored spherical molds. A key contribution of this work is the implementation of a physics-based Law of Cosines separation method to isolate the specular and diffuse reflection components. This allows for the use of strictly diffuse Fresnel equations, which helps to minimize potential zenith or azimuth ambiguities. The surface normals were constructed using the AoP and zenith calculated from normalized Stokes vectors. The resulting normals were then validated by comparing them with the ground truth normals calculated using a custom MATLAB application. The Mean Absolute Error (MAE) and Angular AoP Error were used to quantify reconstruction accuracy. Surface normals still tend to exhibit an azimuth ambiguity, which led to the creation of surrogate surface normals that apply a 2*AoP wrap in order to get rid of this ambiguity. These surrogate normals are no longer true surface normals, but can be used advantageously with pose estimation and classification.

Keywords

Electrical Engineering, Engineering, Physics

Rights Statement

Copyright 2026, author

Link to Full Text

Share

COinS