A Comparative Analysis of Deep Learning Approaches for Visual Perception Benchmarks in Ship Navigation

Abstract

The establishment of a reliable benchmark for evaluating model performance is critical for advancing deep learning (DL), including its application in the recognition of the ship navigation environment. Despite the steady progress being made in object detection models across various tasks, maritime navigation presents unique challenges, such as long distances, miscellaneous objects, wide perception scales, and local conditions and features of water areas. Therefore, the improvement of DL approaches for this domain remains a significant challenge. Using a widely applicable offshore image dataset from the ship bridge, we evaluated the performance of the state-of-the-art object detection model from three perspectives: average precision, multiscale feature calculation, and intersection-over-union design, and explored the factors that may affect the model performance evaluation benchmark from the perspective of data quality, scale calculation, feature quantification, and object association. Our experiments have demonstrated that, in the context of object detection tasks within complex water surface traffic scenes, comprehensive model performance evaluation benchmarks are essential. Such benchmarks must incorporate multiple dimensions of the model.

Keywords

Long-range perception / Visual navigation / Dataset / Multiscale detection / Vision benchmark

RIGHTS & PERMISSIONS

Harbin Engineering University and Springer-Verlag GmbH Germany, part of Springer Nature