Home // COCE 2025, The Second International Conference on Technologies for Marine and Coastal Ecosystems // View article

Semi-Supervised Object Detection for Marine Monitoring using Temporal Information

Authors:
Viljar Holm Elvevoll
Kim Tallaksen Halvorsen
Ketil Malde

Keywords: image classifcation; machine learning; species recognition; semi-supervised learning.

Abstract:
Accurate and sustainable monitoring of marine biodiversity is crucial for effective fisheries management and conservation. Traditional fish population assessments, relying on manual annotation and invasive techniques, are labor-intensive and potentially harmful to marine ecosystems. This work presents a Semi-Supervised Learning (SSL) approach that leverages extensive unlabeled underwater video data to significantly enhance object detection performance for fish species. By integrating the YOLOv8 object detector with Multi-Object Tracking (MOT) algorithms, specifically ByteTrack, a novel methodology is proposed to generate high-quality pseudolabels from temporal sequences. Iterative training incorporating these pseudolabels consistently improved model precision and recall, with the best-performing approach (ByteTrack with an extrapolated heuristic) demonstrating average precision of 90%, recall of 70%, mAP50 of 74%, and mAP50-95 of 59%. Notably, scores improved substantially over the baseline supervised model on all metrics. These results underscore the potential of temporally informed pseudolabeling in enhancing fish detection accuracy and robustness, reducing reliance on manual annotations and supporting sustainable marine monitoring practices

Pages: 27 to 31

Copyright: Copyright (c) IARIA, 2025

Publication date: October 26, 2025

Published in: conference

ISBN: 978-1-68558-329-3

Location: Barcelona, Spain

Dates: from October 26, 2025 to October 30, 2025