Model Card: YOLOv11n Object Detection Model for Mothra Hydrothermal Vent Site

Overview

This YOLOv11n object detection model is designed for automated marine species identification at the Mothra hydrothermal vent site. The model has been trained on images captured by Ocean Networks Canada and is tailored to detect various marine organisms across 12 classes.

Dataset Details

Number of Images: 3,495
Missing Annotations: 0
Null Examples: 341
Total Number of Annotations: 78,001
(Average of 22.3 annotations per image)
Number of Classes: 12
Image Resolutions:
- Average Image Size: 0.70 megapixels
- Range: 0.70 MP to 2.07 MP
- Median Resolution: 1088 x 640 (wide orientation)

Class Distribution

Class	Annotations
Gastropod	59,653
Scaleworm	14,990
Seaspider	1,980
Squatlobster	634
Spidercrab	274
Eelpout	227
Snailfish	130
Cusk-eel	42
Rattail	28
Amphipod	23
Other Classes	Annotations not detailed

Note: While 12 classes are mentioned, detailed annotation counts are provided for 10 classes.

Model Architecture

The model builds on the YOLO framework with a lightweight configuration optimized for fast inference and efficient training. Specific architectural details (such as layer configuration and parameters) can be added as needed based on further documentation.

Training and Validation Splits

The dataset has been partitioned into training, validation, and test sets. Detailed split ratios and performance metrics (precision, recall, mAP, etc.) can be incorporated here as they become available.

Data Collection and Annotation Process

Data Source: Images captured by Ocean Networks Canada from the Mothra hydrothermal vent site.
Data Acquisition: Downloaded using the oncvideo Python package.
Image Subsampling: For each video, the sharpest frame was selected using a Laplacian-based method.
Annotation: Annotations were generated following a process similar to the one described in this resource (I wrote and own this resource).

Intended Use

This model is intended for:

Marine Research: Facilitating ecological studies and biodiversity assessments in deep-sea environments.
Automated Monitoring: Supporting the identification and tracking of marine species at hydrothermal vent sites.
Environmental Management: Assisting in monitoring the health and dynamics of underwater ecosystems.

Limitations

Class Imbalance: A significant imbalance exists in the dataset (e.g., gastropods and scaleworms dominate the annotation counts), which may affect performance on underrepresented classes.
Specificity: Trained exclusively on data from the Mothra hydrothermal vent site, the model may require additional tuning for other underwater settings.
Resolution Variability: Variability in image resolution (0.70 MP to 2.07 MP) could impact detection accuracy under different conditions.

Ethical Considerations

Environmental Impact: The model is developed for research purposes and should be applied in a manner that respects the natural marine environment and wildlife. NO OIL OR GAS COMPANY USE. USE OF THIS SOFTWARE BY COMMERCIAL OR STATE ACTING OIL AND GAS COMPANIES IS PROHIBITED TO EXTENT OF LICENSING

Data Accessibility

The dataset is publicly accessible via Ocean Networks Canada.

Attribution

Data was gathered from: Ocean Networks Canada Society. 2020. Mothra Video Camera Deployed 2020-09-10. Ocean Networks Canada Society. https://doi.org/10.34943/c19da29d-735b-42e2-b932-865a9dd857f8 Data was downloaded using the oncvideo Python package, images were subsampled by capturing the sharpest frame from each video using a Laplacian-based method, and annotations were generated following a process similar to this resource (check it out to replicate on your own dataset!).

atticus-carter
/

YOLOV11_ONC_Mothra