A data-centric framework for combating domain shift in underwater object detection with image enhancement

Abstract

Underwater object detection has numerous applications in protecting, exploring, and exploiting aquatic environments. However, underwater environments pose a unique set of challenges for object detection including variable turbidity, colour casts, and light conditions. These phenomena represent a domain shift and need to be accounted for during design and evaluation of underwater object detection models. Although methods for underwater object detection have been extensively studied, most proposed approaches do not address challenges of domain shift inherent to aquatic environments. In this work we propose a data-centric framework for combating domain shift in underwater object detection with image enhancement. We show that there is a significant gap in accuracy of popular object detectors when tested for their ability to generalize to new aquatic domains. We used our framework to compare 14 image processing and enhancement methods in their efficacy to improve underwater domain generalization using three diverse real-world aquatic datasets and two widely used object detection algorithms. Using an independent test set, our approach superseded the mean average precision performance of existing model-centric approaches by 1.7–8.0 percentage points. In summary, the proposed framework demonstrated a significant contribution of image enhancement to underwater domain generalization.