FathomNet/trash-detector

Trash Detector

Model Details

• Trained by researchers at the Monterey Bay Aquarium Research Institute (MBARI).
• Ultralytics YOLOv8x
• Object detection model
• Classes included in this detection model:
  • trash
  • eel
  • rov
  • starfish
  • fish
  • crab
  • plant
  • animal_misc
  • shells
  • bird
  • shark
  • jellyfish
  • ray

Intended Use

• Post-process video and images collected by marine researchers
• This model should do a reasonable job detecting marine debris in a variety of habitats, depths, and lighting conditions.
• Can be used to build a localized set of training images, when neither training data nor a model exists for the imagery being analyzed.

Factors

• Distribution shifts related to sampling platform, camera parameters, illumination, and deployment environment are expected to impact model performance

Metrics

TODO

Training and Evaluation Data

• Fine-tuned to detect 13 classes using training data combined from the following sources:
  1. MBARI/FathomNet
  2. trash-can: https://conservancy.umn.edu/handle/11299/214865
  3. deep plastic: https://github.com/gautamtata/DeepPlastic
  4. taco-dataset: https://tacodataset.org/
  5. ocean agency image bank: https://www.theoceanagency.org/search-result?s=trash
  6. Trash-ICRA19: https://conservancy.umn.edu/handle/11299/214366
  7. roboflow aquarium dataset
  8. roboflow Underwater Trash Detection.v5-dataset_v3
• A compiled list of trash training data sets is here: https://github.com/AgaMiko/waste-datasets-review

Deployment

1. Clone this repository
2. In an environment with the ultralytics Python package installed, run:

yolo predict model=trash_mbari_09072023_640imgsz_50epochs_yolov8.pt