MassID45

This repo contains the preprocessing and training/inference code for MassID45: A multi-modal dataset for insect biodiversity with imagery and DNA at the trap and individual level.

Installation

Requirements for downloading and preprocessing the data can be found in the image_preprocessing folder. We recommend creating a dedicated Python environment for image_preprocessing.

The CutLER, Mask-RCNN, Mask2Former, MaskDINO, and Grounded-SAM-2 submodules each contain their own installation instructions (see the INSTALL.md files in each submodule). We recommend creating separate Python or Conda environments for each submodule.

Usage

Annotation utilities

Code to run the watershed algorithm and crop the bulk images into annotator patches is contained in the arth-imrec submodule.

Downloading bulk imagery data

The bulk images can be downloaded from Google Drive by running image_preprocessing/download_data.py. Scripts for downloading the ENA sequence data can be found in the image_preprocessing/ENA_sequence_scripts folder.

Pre-processing bulk images for machine learning

The image_preprocessing folder contains utility scripts for downloading and assembling the bulk image data. The data pipeline is as follows:

1. Download the bulk images download_data.py
2. Assemble the annotated patches into the full bulk images assemble_annotations.py
3. Crop the bulk images to exclude excess areas crop_bulk_imgs.py
4. Divide the bulk images into tiles for training tile_imgs.py
5. And lastly, postprocess the sliced annotations postprocess_dataset.py The entire preprocessing pipeline can be run via

sbatch pipeline.sh

Training and Inference

The CutLER, Mask-RCNN, Mask2Former, MaskDINO, and Grounded-SAM-2 submodules each have their own usage instructions (see the MassID45 Instructions section in each submodule's README.md). We ran our experiments using 4 RTX6000 GPUs for training, and 1 RTX6000 GPU for testing.

Model Checkpoints

Model checkpoints can be downloaded at Zenodo.