Real-Time Ship Segmentation in Maritime Surveillance Videos Using Automatically Annotated Synthetic Datasets.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Maritime Surveillance

Background:

• Real-time object detection and segmentation are crucial for maritime surveillance.
• Existing methods often lack accuracy or require extensive manual annotation.
• Unmanned aerial vehicles (UAVs) offer a flexible platform for maritime monitoring.

Purpose of the Study:

• To develop a real-time ship instance segmentation system for UAV-based maritime surveillance.
• To address the lack of suitable annotated maritime video datasets.
• To improve segmentation accuracy and temporal stability in challenging conditions.

Main Methods:

• A two-stage system combining an instance segmentation network with a 3D Conditional Random Field (CRF).
• Development of a synthetic maritime surveillance dataset (MarSyn) with automatic labeling.
• Training and validation using both synthetic and publicly available aerial datasets.

Main Results:

• The proposed system achieves real-time performance for ship instance segmentation.
• The 3D CRF significantly enhances segmentation accuracy by leveraging temporal information.
• The MarSyn dataset facilitates robust model training and validation.

Conclusions:

• The developed system offers a reliable solution for real-time maritime surveillance.
• Synthetic data generation is an effective strategy to overcome annotation limitations.
• The approach demonstrates robustness in handling missing frames and segmentation errors.