Comparing CNN-Based Object Detectors on Two Novel Maritime Datasets

Vessel detection studies conducted on inshore and offshoremaritime images are scarce, due to a limited availability ofdomain-specific datasets. We addressed this need collectingtwo datasets in the Finnish Archipelago. They consist of images of maritime vessels engaged in various operating scenarios, climatic conditions and lighting environments. Vesselinstances were precisely annotated in both datasets. We evaluated the out-of-the-box performance of three state-of-the-artCNN-based object detection algorithms (Faster R-CNN [1],R-FCN [2] and SSD [3]) on these datasets and compared themin terms of accuracy and run-time. The algorithms were previously trained on the COCO dataset [4]. We explore theirperformance based on different feature extractors. Furthermore, we investigate the effect of the object size on the algorithm performance. For this purpose, we group all objects ineach image into three categories (small, medium and large)according to the number of occupied pixels in the annotatedbounding box. Experiments show that Faster R-CNN withResNet101 as feature extractor outperforms the other algorithms.