Trained YOLO Models for Underwater Fish Detection

A YOLO v8 model achieved a mean Average Precision (mAP) of 0.68 on the Eyesea optical dataset and up to 0.65 mAP on unseen sonar data. The submission includes trained model weights, code for five experiments, and documentation for replicating fish detection research. The project aims to improve monitoring around marine energy facilities using optical and sonar imagery.

Use Cases

• Apply the provided YOLO model weights (.pt files) for object detection on new underwater optical or sonar imagery.
• Replicate the hyper-image analysis experiment using the provided Python and shell scripts to stack consecutive frames for enhanced detection.
• Use the sample data structures to format raw Eyesea or similar sonar datasets for training or inference with the included models.
• Leverage the cross-domain adaptability demonstrated between the Eyesea optical dataset and Alaska Fish and Games sonar images for transfer learning.

Strengths

• Model performance is quantified with a best mAP of 0.68 on the Eyesea dataset, surpassing a cited YOLO v3 benchmark.
• Includes five complete experiment packages with Python, shell, and YAML configuration files for full reproducibility.
• Demonstrates cross-domain application with a reported mAP of up to 0.65 on an unseen sonar domain using a hyper-image approach.

Limitations

• The underlying raw Eyesea and sonar image datasets are not included; users must download and format them separately using provided instructions.
• Model performance metrics are specific to the tested optical and sonar domains; applicability to other underwater imaging conditions is unverified.
• The submission is a model and code repository; there is no tabular data with defined rows or columns for traditional analysis.

Provenance

Source

Department of Energy.

Collection Method

YOLO models were trained using the Eyesea optical dataset and sonar images from Alaska Fish and Games via supervised methods.

Freshness

Last updated in May 2025.

Geography

Imagery includes the Eyesea dataset and sonar data from Alaska; intended for use around marine energy facilities.