FathomNet/fathomnet2023-comp-baseline

FathomNet2023 Baseline Model

Model Details

• Trained by researchers at Monterey Bay Aquarium Research Institute (MBARI) as a baseline for the FathomNet2023 Competition presented with the Fine Grained Visual Categorization workshop at CVPR 2023.
• Ultralytics YOLOv8.0.117
• Object detection
• Fine tuned yolov8m to detect 290 fine grained taxonmic categories of benthic animals in the Greater Monterey Bay Area off the coast of Central California.

Intended Use

• Make detections on images collect on the sea floor in the Monterey Bay Area.

Factors

• Distribution shifts related to sampling platform, camera parameters, illumination, and deployment environment are expected to impact model performance.
• Evaluation was performed on an IID subset of available training data.
• Data to test out of distribution performance can be found on the competition Kaggle page.

Metrics

• Precision-Recall curve and per class accuracy were evaluated at test time.
• [email protected] = 0.33515
• Performance is quite variable depending on the target organism even when testing on in-distribution data.
• Identified out-of-sample images with a binary metric, returning ROC ~= 0.7.

Training and Evaluation Data

• Training data is the FathomNet2023 competition split and internal MBARI data
• Class labels have a long tail and localizations occur throughout the frame.

Deployment

In an environment running YOLOv8:

python classify/predict.py --weights fathomnet23-comp-baseline.pt --data data/images/