TL;DR
This paper introduces UOW-Vessel, the largest publicly available dataset for fine-grained vessel detection and segmentation in high-resolution satellite imagery, featuring precise polygon annotations. They also provide a comprehensive benchmark of state-of-the-art instance segmentation models, highlighting YOLOv8 as a top performer.
What Problem Does This Paper Address?
Maritime surveillance relies heavily on identifying vessels from satellite imagery. While existing datasets exist, they often lack the detail needed for fine-grained classification (e.g., distinguishing between different types of military vessels) or provide only coarse bounding box annotations. This limits the application of advanced techniques like instance segmentation, which requires precise pixel-level masks to accurately capture vessel shapes, occlusions, and orientations in complex real-world scenes. The authors aim to bridge this gap by creating a more comprehensive dataset and evaluating current instance segmentation methods on it.
Methods
The paper focuses on evaluating existing instance segmentation methods. They categorize these methods into three groups:
- Two-stage methods: These first propose candidate regions and then refine them for detection and segmentation. Mask R-CNN is a prime example, building upon Faster R-CNN by adding a mask prediction branch.
- Multi-stage methods: Cascade Mask R-CNN extends the cascade approach for instance segmentation by incorporating mask prediction at various stages of the detection pipeline.
- Single-stage methods: These perform detection and segmentation in a single network pass, aiming for faster inference. Examples include SOLOv2, CondInst, and various YOLO architectures (YOLOv5, YOLOv7, YOLOv8), which have been adapted for instance segmentation by adding a dedicated segmentation head.
The authors conducted extensive experiments to compare these methods on their new dataset, evaluating them across different annotation complexity levels.
Datasets
The core contribution is the UOW-Vessel dataset:
- Size: 3,500 high-resolution optical satellite images.
- Coverage: Collected from 14 countries across 4 continents.
- Instances: 35,598 annotated vessel instances.
- Categories: 10 distinct vessel categories, including specific military types like Aircraft Carrier, Cruiser, Destroyer, Frigate, Patrol, and Corvette, as well as Civilian vessels.
- Annotations: Precise polygon annotations for each vessel instance, offering a significant advantage over bounding boxes for capturing intricate details and handling occlusions.
- Comparison to Existing Datasets: UOW-Vessel is highlighted as the largest satellite image dataset for vessel recognition to date, offering more detailed annotations than datasets like HRSC2016, ShipRSImageNet, and VHRShips, which primarily provide bounding boxes. It also surpasses datasets like FGSC-23 and FGSCR-42 in terms of image complexity and annotation detail.
Key Insights and Ideas
- Polygon Annotations are Crucial: The paper strongly advocates for polygon annotations, demonstrating their necessity for accurately representing complex vessel shapes and handling real-world scenarios like occlusions and arbitrary orientations, which are critical for fine-grained recognition.
- Instance Segmentation is Underexplored: Despite the growing interest in deep learning for vessel recognition, instance segmentation remains a less investigated area compared to object detection. UOW-Vessel aims to spur research in this direction.
- Class Imbalance is Natural: The dataset exhibits a natural class imbalance, with "Civilian" vessels being the most numerous. This reflects real-world distributions and presents a challenge for model training.
- YOLO Architectures Lead: The experimental results consistently show that YOLO-based models (YOLOv5, YOLOv7, YOLOv8) achieve superior performance in terms of both detection and segmentation accuracy, often outperforming traditional two-stage and multi-stage methods.
- Higher Resolution Matters: Increasing input image resolution significantly improves detection and segmentation accuracy, though it comes with a trade-off in inference speed.
Key Results
- UOW-Vessel Dataset: The largest satellite image dataset for vessel recognition with fine-grained, polygon annotations.
- Model Performance: YOLOv8-Large achieved the highest overall mAP across different annotation levels, demonstrating strong detection and segmentation capabilities.
- YOLO vs. Others: YOLOv5, YOLOv7, and YOLOv8 generally outperformed Mask R-CNN and Cascade Mask R-CNN in mAP metrics. For instance, YOLOv8 achieved an mAP of 80.0 at Level 2, compared to 53.4 for Mask R-CNN.
- Inference Speed: YOLO models also offered competitive inference speeds, with YOLOv8 achieving 35.5 FPS, meeting real-time requirements.
- Challenging Conditions: Performance degrades in poor illumination and cloud/haze conditions, indicating a need for more robust models.
- Resolution Impact: Increasing image resolution from 480x480 to 1440x1440 pixels improved mAP by up to 19.1%, highlighting the importance of high-resolution input.
Limitations and Future Work
- Class Imbalance: The dataset exhibits a natural class imbalance, which can affect model performance on underrepresented classes.
- Inter-Class Similarity: Misclassification between visually similar vessel categories remains a challenge, even for top-performing models.
- Challenging Conditions: Performance drops in adverse weather and lighting conditions suggest a need for more robust feature extraction or domain adaptation techniques.
- Future Work: The authors expect UOW-Vessel to facilitate the development of new algorithms for vessel detection and segmentation. Further research could focus on improving robustness to challenging conditions, addressing class imbalance, and exploring more efficient architectures for real-time applications.
Summary updated