Road scene object detection using pre-trained RGB neural networks on linear Stokes images

Neural networks trained on RGB and monochromatic images are tested on images augmented by polarimetry for recognition of road-based objects. The goal of this work is to understand the scene conditions for which object detection and recognition can be improved by linear Stokes measurements. Shadows, windows, low albedo, and other object features which reduce RGB image contrast also decrease neural network detection performance. This work demonstrates specific cases for which linear Stokes images increase image contrast and therefore increase object detection by a neural network. Linear Stokes videos for five difference scenes are collected at three times of day and two driving directions. Although limited in scope, this work demonstrates some enhancement to object detection by adding polarimetry to neural networks trained on RGB images.