Cross-GOOSe

Crossing Gradient-flows for Overlapping Objects Segmentation

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

📄 Citation

@inproceedings{mabon2026crossgoose,
  author = {Mabon, J. and Olivo-Marin, J. C.},
  title = {Improving Gradient Flow Methods for Instance Segmentation of Crossing Objects},
  booktitle = {2026 IEEE International Symposium on Biomedical Imaging},
  address = {London, United Kingdom},
  month = {April},
  year = {2026}
}

📄 Paper on HAL

👩🏻‍💻 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

🌟 Changelog

v1.3.5

• Trajectory-based training: new TrajectorySampler and train_on_trajectories option to learn from full point trajectories instead of just (u0, ut) pairs
• Time-error weighting: time_error_weighting option to balance loss contribution across time steps
• CPLike flow function: new FlowFunction variant that uses only et embedding (CellPose-like)
• Improved GridFlow:
  • new from_labels() method for direct label array loading
  • keep_largest_component() to handle non-contiguous masks
  • threaded query_multiple_labels_threaded() for faster queries
• 2-channel image support: model now handles both grayscale and multi-channel images (grayscale flag in dataset)
• Enhanced point sampling: RandomOnCellV2 with improved random sampling within instances
• Augmentation updates: fixed patching and transforms for multi-channel images
• New configs: multiple versioned configs (1.3.0 through 1.3.5) for reproducibility

v1.2.1

• Shared embedding architecture: new shared_embedding option to use the same embedding for both u0 and ut, reducing model parameters
• Embedding visualization: added notebook and utilities to view learned embeddings
• Model configuration: added model_sharedemb.yaml config for shared embedding training

v1.1.1

• GridFlow optimization: major refactor of flow computation with BatchGridFlow for faster precomputing
• Data pipeline refactor: moved sampling to dataloader, reorganized crossgoose/data/ module structure
• Threading improvements: added threaded point sampling for faster data loading
• Configuration updates: new model config system (model2.yaml), updated defaults
• Bug fixes: fixed normalization vector, gridflow augmentation, and removed CUDA flow compute