An Image-Based Benchmark Dataset and a Novel Object Detector for Water Surface Object Detection.

Area of Science:

• Computer Vision
• Autonomous Systems
• Machine Learning

Background:

• Water surface object detection is crucial for autonomous driving and specialized vision applications.
• Existing datasets lack scale and scenario specificity, hindering algorithm development.

Purpose of the Study:

• To introduce a large-scale, high-quality benchmark dataset for water surface object detection.
• To propose and evaluate a novel baseline object detection network, CRB-Net.

Main Methods:

• The creation of the Water Surface Object Detection Dataset (WSODD) with 7,467 images and 21,911 instances across 14 categories.
• Development of CRB-Net, a straightforward yet effective object detection architecture.
• Comparative analysis of CRB-Net against 16 state-of-the-art methods.

Main Results:

• WSODD encompasses diverse water environments, climates, and times, focusing on specific scenarios.
• CRB-Net demonstrated superior detection precision compared to all 16 evaluated state-of-the-art methods.
• Cross-dataset validation confirmed WSODD's superiority over existing datasets and CRB-Net's excellent adaptability.

Conclusions:

• The proposed WSODD dataset significantly enhances the evaluation of water surface object detection algorithms.
• CRB-Net establishes a strong baseline, outperforming existing methods and showing excellent adaptability.
• This work accelerates progress in autonomous driving and water surface vision.