A Benchmark for Deep Learning Based Object Detection in Maritime Environments

Abstract

Object detection in maritime environments is a rather unpopular topic in the field of computer vision. In contrast to object detection for automotive applications, no sufficiently comprehensive public benchmark exists. In this paper, we propose a benchmark that is based on the Singapore Maritime Dataset (SMD). This dataset provides Visual-Optical (VIS) and Near Infrared (NIR) videos along with annotations for object detection and tracking. We analyze the utilization of deep learning techniques and therefore evaluate two state-of-the-art object detection approaches for their applicability in the maritime domain: Faster R-CNN and Mask R-CNN. To train the Mask R-CNN including the instance segmentation branch, a novel algorithm for automated generation of instance segmentation labels is introduced. The obtained results show that the SMD is sufficient to be used for domain adaptation. The highest f-score is achieved with a fine-tuned Mask R-CNN. This is a benchmark that encourages reproducibility and comparability for object detection in maritime environments.