Detection and Metric Learning from Pseudo-labels for Unsupervised Maritime Vessel Identification

This thesis introduces an unsupervised framework that exploits model agreement, temporal coherence, and geometric priors from Automatic Identification System (AIS) geolocation messages to produce pseudo-labels of maritime vessels from raw surveillance streams suitable for vessel detection and vessel re-identification. Data sources include daylight and thermal camera frames, AIS data, and camera calibration information (field of view, azimuth, elevation, and intrinsic/extrinsic parameters). The first contribution is a pseudo-label generation pipeline for vessel detection, comprising a majority-voting ensemble of strong CNN and ViT detectors, a SAM3-based temporally consistent tracker, and ELoFTR-based keypoint matching for cross-modal detection transfer with homographies. A lightweight You Only Look Once detector is trained on the generated annotations. The second contribution is a pseudo-label generation pipeline for vessel re-identification, cross-referencing AIS geodetic positions with YOLO detections via perspective projection. Finally, a metric-learning recipe is presented and evaluated with both classical ResNets and the latest DINOv3 backbones. Hence, the motivation behind this work is to eventually allow the transformation of maritime surveillance streams into practical datasets for downstream tasks. Experimental validation confirms the effectiveness of the proposed framework across both detection and re-identification tasks, with the fine-tuned YOLO26l detector reaching 0.951 mAP@50 and 0.916 F1 score, while the best Re-ID setting based on DINOv3 ViT achieves 0.996 centroid mAP.