Dataset: OzFish
OzFish is a collection of ~80k fish crops, ~45k bounding box annotations derived from Baited Remote Underwater Video Stations (BRUVS) and comprised of 70 families, 200 genera and 507 species of fish. This dataset is completely open and free to use for advancing machine learning for the classification of fish from underwater imagery. This dataset has been developed as part of the Australian Research Data Commons Data Discoveries program with the aim to further advance research into machine learning for the automated detection of fish from video. There are also 16 species with more than 1000 detections and 620 different species with 80983 detections from 64385 images from 1013 videos. The images also have bounding boxes with JSON metadata.
Data Access link: https://apps.aims.gov.au/metadata/view/38c829d4-6b6d-44a1-9476-f9b0955ce0b8
Model: YOLOv7
The model was trained using the default hyperparameters of YOLOv7 model. No pre-processing was done before training. The model was trained for 500 epochs and the model with the best validation mAP is uploaded to the repo.