MID: A Comprehensive Shore-Based Dataset for Multi-Scale Dense Ship Occlusion and Interaction Scenarios

Yugang Chang, Hongyu Chen, Fei Wang, Chengcheng Chen, Weiming Zeng · 2024

Read MID: A Comprehensive Shore-Based Dataset for Multi-Scale Dense Ship Occlusion and Interaction Scenarios by Yugang Chang, Hongyu Chen, Fei Wang, Chengcheng Chen, Weiming Zeng (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-05871.txt.

Introduction

Busy ports need cameras that can spot ships even when vessels overlap, cluster together, or hide in fog and rain. Yugang Chang and four co-authors present MID in this 2024 paper, a shore-based photo set built for that hard task. The images use tilted boxes fitted to each ship rather than plain upright boxes, which suit long hulls at odd angles.

The set holds 5,673 images with 135,884 labeled ships, drawn from video at forty-three water areas in varied weather and light. Scenes cover ship meetings, docking, small craft in crowds, and partly hidden vessels. The team tested ten detection methods on the set and compared how each coped with overlap and dense groups. They argue the set fills gaps left by older ship data and can sharpen port watch systems.

Catalog record

CREATORYugang Chang, Hongyu Chen, Fei Wang, Chengcheng Chen, Weiming Zeng
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-05871

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MID: A Comprehensive Shore-Based Dataset for Multi-Scale Dense Ship Occlusion and Interaction Scenarios Yugang Chang, Hongyu Chen, Fei Wang, Chengcheng Chen, Weiming Zeng Abstract This paper introduces a maritime ship navigation behavior dataset, referred to as MID, which is annotated using Oriented Bounding Boxes (OBB) to address the challenges of ship target detection in complex real-world maritime environments. MID currently contains 5,673 images with a total of 135,884 finely annotated target instances, providing a robust foundation for both supervised and semi-supervised learning tasks. The dataset encompasses a variety of maritime scenarios, including ship encounters in diverse weather conditions, docking maneuvers, small target aggregation, and partial occlusions. These scenarios contribute to filling the gap in existing datasets (e.g., HRSID, SSDD, NWPU-10) by addressing the lack of data on complex situations such as occlusion and dense interactions, which are particularly relevant in the context of advanced technologies such as autonomous ships. All images in MID are sourced from high-definition video clips of real-world maritime navigation across 43 distinct navigable water areas. The dataset is further enriched through the inclusion of multiple weather and lighting conditions (e.g., rainy, cloudy, foggy days) and manually supplemented annotations, enhancing its diversity and ensuring that models trained on MID are better equipped to handle varying natural environments. This diversity also ensures its applicability to the real-world demands of busy ports and dense maritime regions. Using MID, we conducted a comprehensive evaluation of 10 detection algorithms, presenting a detailed analysis and comparison of their performance. This study includes: 1) an in-depth analysis of the dataset, 2) a showcase of detection results from various models, and 3) a comparison of baseline algorithms, with particular focus on their ability to detect dense occlusions. The results suggest that MID holds significant potential for advancing research in intelligent maritime traffic monitoring systems and autonomous vessel navigation, with important implications for future applications in maritime environment and safety situational awareness. Our dataset will be released at https://github.com/VirtualNew/MID_DataSet . \FAILED \FAILED I Introduction In the context of the global development of smart ships and maritime port traffic management, the accuracy of target detection, tracking, and trajectory prediction for ships in complex scenarios is crucial for ensuring port safety, enhancing logistics efficiency, and reducing maritime accidents [ 1 ] . With the rapid growth of the global shipping industry, ports have increasingly become hubs of international trade [ 2 ] . Therefore, ensuring the safe navigation and efficient operation of ships entering and leaving ports in complex environments is especially important. Modern ports are typically equipped with numerous surveillance cameras to monitor the entry, exit, and activities of ships in real time [ 3 ] . However, the complexity of port environments characterized by a high number of ships, overlapping navigation paths, and diverse ship sizes and types poses significant challenges for ship detection and tracking [ 4 ] . In densely populated areas, smaller ships are particularly prone to being obscured or confused with other ships, further complicating detection and tracking tasks [ 5 ] . Addressing these issues is not only vital for the operational efficiency of ports but also plays a key role in reducing the incidence of maritime accidents. At the same time, with the advancement of autonomous ship technology, intelligent navigation has become a research hotspot. The intelligent navigation of autonomous ships relies on three core technologies: environmental and safety situational awareness, route planning and collision avoidance decision-making, and ship motion control. Among these, environmental and safety situational awareness [ 6 ] is a prerequisite for autonomous navigation, particularly the reliable identification of the relative motion of external ship targets, which directly impacts navigational safety. In the early stages of autonomous ship development, navigational safety should be regarded as the primary objective of perception systems [ 7 ] . However, the complexity of navigation environments and behaviors makes this task more challenging. When constructing datasets for ship target detection, it is essential to consider multiple influencing factors in complex navigation scenarios. First, the diversity of all-weather external environments, including different weather conditions (such as sunny, cloudy, rainy, and foggy days) and lighting conditions (daytime and nighttime), as well as background noise from sea spray and cloud shadows [ 8 ] . Second, the complex navigational behaviors between ships, such as crossing, overtaking, and meeting behaviors that occur in open waters and busy navigation areas, which may lead to occlusion and pixel overlap in images [ 9 ] . Finally, the variations in shape, size, and tonnage of different ships also affect the detection model, as differences in ship size result in significant variations in the pixel distribution of targets within images [ 10 ] . Therefore, effective perception and processing of these complex factors are crucial for achieving safe and efficient navigation in autonomous ship systems. Whether in traditional port surveillance or in the intelligent navigation of autonomous ships, the core challenges of ship detection, tracking, and identification are similar. Both require accurate recognition of dense and occluded targets in complex environments. This necessitates not only advanced target detection technologies but also the development of robust datasets that can comprehensively cover the diverse navigation environments and ship behaviors. Despite the widespread use of automatic identification systems (AIS) and radar for ship monitoring, they have significant limitations [ 11 ] . AIS relies on ships actively transmitting signals, which means it cannot monitor ships that have not turned on their AIS or whose signals are disrupted [ 12 ] . Although radar is effective in low visibility conditions, it struggles with target overlap and interference in dense environments, making it challenging to accurately distinguish complex interactions between ships [ 13 ] . Additionally, radar has lower resolution and provides less detailed and intuitive information about ship targets, limiting its effectiveness for detecting smaller ships [ 14 ] . In contrast, optical images, with their high resolution, low cost, and rich visual information, have emerged as an ideal data source for ship target detection, particularly for real-time monitoring at medium to short distances [ 15 ] . Compared to infrared thermal cameras, low-light night vision devices, and millimeter-wave radar, optical images offer more precise target detection in congested navigation areas and are relatively less affected by environmental and weather conditions [ 16 ] . They can clearly capture the shape and position of ships, excelling in complex and dense scenarios. When combined with deep learning techniques, optical data can effectively address detection challenges posed by targets of various sizes, occlusion, and intricate interaction scenarios, providing robust support for the environmental perception of autonomous ships. Moreover, existing maritime datasets often overlook the impact of marine weather conditions, such as rain, fog, overcast skies, and sunny days, as well as issues like image blurriness caused by weather [ 17 ] . These conditions impose higher demands on ship detection and tracking, increasing the challenges posed by low visibility, varying lighting, and noise interference. Therefore, for small target detection and tracking in scenarios characterized by diverse sizes, dense occlusion, complex interactions, and varying marine weather conditions, optical data offers significant advantages and holds considerable research and application value. Currently, visual tasks in dense maritime scenarios—such as overtaking, crossing, and meeting—are increasingly becoming an important research

Read the complete text