Advancing Underwater Object Detection: A Comparative Study of YOLOv5s, YOLOv8s, and YOLOv11s with Bias Analysis

Abstract

Environmental distortions and dataset-induced bias are challenges for underwater object detection due to the impact on model generalization. This research paper focuses on bias in the RUOD dataset and compares the performance of YOLOv5s, YOLOv8s, and YOLOv11s with 9,800 training images. Precision, Recall, and mean Average Precision (mAP) were used to perform a comprehensive analysis. The comparative results demonstrate that YOLOv11s has the highest performance when evaluated by mAP@0.5 (0.8721), followed by YOLOv8s and YOLOv5s, with slight differences among the models. The bias analysis of the training and validation sets showed a significant class imbalance, with the dominant classes being fish (17.5%) and echinus (15.15%), and the underrepresented classes being turtle (5.44%) and jellyfish (3.62%). Class-wise evaluation, however, suggests that the accuracy of detecting classes does not directly relate to the frequency of the classes, as there are several classes with low representation that are detected with high accuracy. The results show that there is a complex interplay between the bias in the datasets and the performance of the models, and that architectural improvements are not enough to entirely compensate for the bias effects. This research work gives insights into bias-aware evaluation to enhance underwater object detection systems.