🪿 Cross-GOOSe

Crossing Gradient-flows for Overlapping Objects Segmentation

implementation of Improving Gradient Flow methods for instance segmentation of crossing objects J. Mabon & J.C. Olivo-Marin, submitted to ISBI 2026

Installation

Setup the environment with conda/mamba :

mamba create -f env.yaml -y
mamba activate crossgoose

Training

preparing the data

# get the data
mkdir -p data/BBBC010_v2_images 
wget "https://data.broadinstitute.org/bbbc/BBBC010/BBBC010_v2_images.zip" -O images.zip
unzip images.zip -d data/BBBC010_v2_images
rm images.zip

wget "https://data.broadinstitute.org/bbbc/BBBC010/BBBC010_v1_foreground_eachworm.zip" -O labels.zip
unzip labels.zip -d data
rm labels.zip

# make a dataset
python main.py make_dataset --config configs/dataset.yaml
# make a dataset with synthetic data
python main.py make_synth_dataset --config configs/synth_dataset.yaml

train

python train.py fit --config configs/model.yaml

Infer/eval

python main.py eval_models --config configs/eval.yaml