ViV-ReID: Bidirectional Structural-Aware Spatial-Temporal Graph Networks on Large-Scale Video-Based Vessel...

Area of Science:

• Maritime surveillance
• Intelligent transportation systems
• Computer vision

Background:

• Vessel re-identification (ReID) is crucial for intelligent maritime transportation.
• Existing research is limited by a lack of large-scale video datasets.
• Video data offers richer information than static images but presents challenges in feature extraction.

Purpose of the Study:

• To introduce the first large-scale dataset for video-based vessel ReID, named ViV-ReID.
• To propose a novel network, the Bidirectional Structural-Aware Spatial-Temporal Graph Network (Bi-SSTN), to address challenges in video-based vessel ReID.
• To establish performance benchmarks for video-based vessel ReID methods.

Main Methods:

• Development of the ViV-ReID dataset: 480 identities, 20 camera views, 7,165 tracklets, 1.14 million frames.
• Proposal of the Bidirectional Structural-Aware Spatial-Temporal Graph Network (Bi-SSTN) utilizing vessel structural priors.
• Extensive experiments to evaluate performance and compare with existing methods.

Main Results:

• ViV-ReID dataset established as a benchmark for video-based vessel ReID.
• Image-based ReID methods show suboptimal performance on video data.
• The proposed Bi-SSTN significantly outperforms state-of-the-art methods on the ViV-ReID dataset.

Conclusions:

• The ViV-ReID dataset facilitates advancements in video-based vessel ReID.
• The Bi-SSTN effectively models vessel-specific spatio-temporal patterns, demonstrating the importance of structural priors.
• This work sets a new benchmark for maritime surveillance using video analysis.