MBARI Benthic Supercategory Object Detector

A RetinaNet model fine tuned from the Detectron2 object detection platform's ResNet backbone to identify 20 benthic supercategories drawn from MBARI ROV image data collected in Monterey Bay off the coast of Central California. The data is drawn from FathomNet and consists of 32779 images that contain a total of 80683 localizations. The model was trained on an 85/15 train/validation split at the image level.

confusion_matrix_norm.png is a plot of classifier performance on an independent test set collected in the same study region from the training data. This in-distribution data set is quite close to the images used to tune the network. The high background false positive and negative rates reflect the challenging environment and potential incompleteness of the original annotations. 

This repository contains:

• model_final.pth - Fine tuned model weights
• train_images.json - Training image FathomNet UUIDs, bounding boxes, and meta data
• val_images.json - Validation image FathomNet UUIDs, bounding boxes, and meta data
• benthic_label_map.json - Dictionary that links lowest taxonomic rank to supercategory
• Detectron2-Inference-MBARI.ipynb - Jupyter Notebook that runs model inference with these weights
• Detectron2_Inference_MBARI_multiscale.ipynb - Jupyter Notebook that test running model inference at two resizing scales to better resolve small targets
• fathomnet_config_v2_1280.yaml - Settings file for Detectron2 interface