Quantifying the Degradation of Optical Algorithms in Increasingly Turbid Mediums
Mitchell Scott, Aaron Marburg · OCEANS 2021: San Diego – Porto · 2021
The quality of stereo and monocular camera data is known to degrade as the density of suspended marine particulate - i.e., turbidity-increase. However, few studies have explored the degree to which common computer vision algorithms degrade in the presence of increasing turbidity. In this paper, we explore the quantitative tie between turbidity and the performance of several common computer vision algorithms. We focus on three computer vision operations: 1) the ability to detect and correspond feature points in a pair of images, 2) the ability to generate an accurate depth map from a stereo pair of images, and 3) the performance of deep learning detectors (Faster-RCNN and RetinaNet) to correctly identify an object in the environment. Additionally, for stereo camera feature detection and correspondence, we explore which common feature extractors perform best and how image enhancement impacts algorithm performance. Our results indicate that stereo algorithm performance does drop significantly in the presence of higher turbidity, as is expected. However, we find that optical data remains usable in low to medium turbidity environments. Additionally, we find that contrast limited adaptive histogram equalization (CLAHE) image enhancement has a positive impact on the ability to generate and correspond feature points, and that the introduction of small quantities of turbid data into a deep learning detector’s training step has a significant positive contribution to network performance.