[Paper Review] GLSD: The Global Large-Scale Ship Database and Baseline Evaluations

Zhenfeng Shao, Jiaming Wang+9 more|arXiv (Cornell University)|Jun 5, 2021

Advanced Neural Network Applications51 references4 citations

TL;DR

This paper introduces GLSD, a large-scale global ship database with 152,576 images and 212,357 instances across 13 detailed ship categories, designed to advance ship detection in real-world scenarios. It proposes a standardized evaluation protocol and benchmarks state-of-the-art object detectors, demonstrating that SyncBN normalization and multi-scale training yield measurable performance gains on challenging, diverse ship imagery with geographic context.

ABSTRACT

In this paper, we introduce a challenging global large-scale ship database (called GLSD), designed specifically for ship detection tasks. The designed GLSD database includes a total of 212,357 annotated instances from 152,576 images. Based on the collected images, we propose 13 ship categories that widely exist in international routes. These categories include Sailing boat, Fishing boat, Passenger ship, Warship, General cargo ship, Container ship, Bulk cargo carrier, Barge, Ore carrier, Speed boat, Canoe, Oil carrier, and Tug. The motivations of developing GLSD include the following: 1) providing a refine and extensive ship detection database that benefits the object detection community, 2) establishing a database with exhaustive labels (bounding boxes and ship class categories) in a uniform classification scheme, and 3) providing a large-scale ship database with geographic information (covering more than 3000 ports and 33 routes) that benefits multi-modal analysis. In addition, we discuss the evaluation protocols corresponding to image characteristics in GLSD and analyze the performance of selected state-of-the-art object detection algorithms on GSLD, aiming to establish baselines for future studies. More information regarding the designed GLSD can be found at https://github.com/jiaming-wang/GLSD.

Motivation & Objective

• To address the lack of large-scale, fine-grained, and geographically annotated ship datasets for training and evaluating object detection models.
• To establish a standardized evaluation protocol tailored to the unique challenges of ship detection, including scale, viewpoint, and environmental variations.
• To provide a baseline performance evaluation for state-of-the-art object detection models on a diverse, globally representative ship dataset.
• To support multi-modal analysis by integrating geographic information, including over 3,000 ports and 33 international shipping routes.
• To mitigate long-tail distribution issues by identifying underrepresented ship types such as 'Tug', 'Canoe', and 'Speed boat' for future data expansion.

Proposed method

• The GLSD database was constructed using images collected from satellite and aerial sources, covering international maritime routes and ports, with rigorous annotation of bounding boxes and 13 distinct ship categories.
• A novel evaluation protocol was designed based on image characteristics such as scale, viewpoint, and environmental conditions to assess model robustness and generalization.
• State-of-the-art object detection models, including PAA, were trained and evaluated on GLSD using standardized metrics like AP, AP50, AP75, and AP_S/M/L to measure performance across object sizes.
• Ablation studies were conducted to evaluate the impact of multi-scale training and normalization strategies (BN, GN, SyncBN) on detection accuracy and convergence speed.
• The route-based version, GLSD_port, integrates geographic metadata to enable multi-modal learning and context-aware detection tasks.
• The dataset is publicly released via GitHub to support reproducibility and community-driven research.

Experimental results

Research questions

• RQ1 How does the performance of state-of-the-art object detectors vary on a large-scale, fine-grained, and geographically annotated ship detection benchmark?
• RQ2 What is the impact of multi-scale training and normalization strategies on detecting small and varied ship instances in diverse imaging conditions?
• RQ3 How do different image characteristics—such as scale, viewpoint, and weather—adversely affect detection performance on real-world maritime data?
• RQ4 To what extent does the inclusion of geographic context (ports and shipping routes) improve the robustness and accuracy of ship detection models?
• RQ5 What are the key limitations of current models on GLSD, particularly regarding long-tail distribution in underrepresented ship categories?

Key findings

• The use of Synchronized Batch Normalization (SyncBN) improved the average precision (AP) by 0.3% compared to Batch Normalization (BN) and 3.6% compared to Group Normalization (GN), indicating its superiority in handling GLSD's diverse data distribution.
• Multi-scale training with image sizes of 1,333×640 and 1,333×800 pixels yielded minimal performance gains, suggesting that the extreme scale variation in GLSD limits the benefit of standard multi-scale augmentation.
• The average precision (AP) of the PAA detector on GLSD was 30.9%, with significantly lower performance on small objects (AP_S = 3.9%), highlighting challenges in detecting tiny ships.
• The AR_S (average recall for small objects) was 34.3% with BN, dropping to 31.6% with GN, indicating that normalization strategy significantly affects small-object detection performance.
• The model achieved 72.5% average recall for large objects (AR_L), suggesting that large ships are more reliably detected, but small and medium-sized ships remain a major challenge.
• The dataset exhibits a pronounced long-tail distribution, with low instance counts for 'Tug', 'Canoe', and 'Speed boat', which are critical for real-world maritime monitoring and require further data collection.