ABOships-PLUS

08.01.2024 Updated the annotations to be the correct ones.

ABOships-PLUS is an improved iteration of the original ABOships dataset. It includes 9,880 images capturing maritime scenes, showcasing various types of maritime objects such as powerboats, ships, sailboats, and stationary objects. Detailed category definitions and images can be found in the associated reference paper. In total, ABOships-PLUS contains 33,227 annotated objects across these categories, including four types of ships.

Several key changes and improvements have been made to ABOships-PLUS:

• Object Size Filtering: In ABOships-PLUS, a filtering process was applied to exclude very small objects, specifically those with an occupied pixel area less than 16^2x16^2 pixels. This filtering ensures that the dataset primarily consists of more discernible maritime objects, contributing to improved data quality.
• Superclass Aggregation: A notable transformation in ABOships-PLUS is the grouping of objects into four superclasses based on their distinct visual characteristics. This superclass aggregation facilitates the use of both transfer learning and learning from scratch, making the dataset more versatile for various machine learning applications.
• Semantic Relevance: The categorization into superclasses in ABOships-PLUS was guided by semantic relevance, with human supervision. The objective was to create more meaningful superclasses, both from a semantic and visual perspective, enhancing the dataset's utility for maritime object detection research.
• Format Transition: A significant change occurred in the data format. ABOships-PLUS adopts the COCO format for object detection: https://cocodataset.org/#format-data. This format transition enhances compatibility with a broader range of machine learning frameworks and tools, in contrast to the original ABOships dataset, which used CSV format.

To create ABOships-PLUS, images were extracted from videos recorded in MPEG format, with a resolution of 720p at 15 frames per second (FPS). An image was extracted every 15 seconds, equivalent to every 225 frames, from videos filmed in the Finnish Archipelago using a camera attached to a moving watercraft known as a waterbus or "vesibussi" in Finnish.

The distribution of labels within ABOships-PLUS is as follows: powerboat (21.8%), ship (46.0%), sailboat (24.2%), and stationary objects (8.1%). These changes aim to enhance the dataset's usability for maritime object detection research and applications.

Reference article: https://doi.org/10.3390/jmse11091638