megafishdetector

Detector for generic "fish" trained on publicly available datasets, currently supporting YOLO-style bounding boxes prediction and training. Can also be used as pre-trained networks for further fine-tuning.

Initial experiments to train a generic MegaFishDetector modelled off of the MegaDetector for land animals (https://github.com/microsoft/CameraTraps/blob/main/megadetector.md)

Currently based on YOLOv5 (https://github.com/ultralytics/yolov5).

This repo contains links to public datasets, code to parse datasets into a common format (currently YOLO darknet only), and a model zoo for people to start with. For instructions to run, see the link above.

Instructions

1. Install Yolov5
2. Download desired network weights
3. Usage (from yolov5 root): python detect.py --imgsz 1280 --conf-thres 0.1 --weights [path/to/megafishdetector_v0_yolov5m_1280p] --source [path/to/video/image folder]

Public Datasets Used in v0:

• AIMs Ozfish
• FathomNet
• VIAME FishTrack
• NOAA Puget Sound Nearshore Fish (2017-2018)
• DeepFish
• NOAA Labelled Fishes in the Wild

To Cite:

paper

@misc{yang2023biological,
      title={Biological Hotspot Mapping in Coral Reefs with Robotic Visual Surveys}, 
      author={Daniel Yang and Levi Cai and Stewart Jamieson and Yogesh Girdhar},
      year={2023},
      eprint={2305.02330},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}

TODO:

• Train larger models
• requirements.txt for things like fathomnet environment
• COCO format output