How junhaolixd/biohub_cell_tracking works
Biohub is a Kaggle cell-tracking solution for 3D time-lapse zebrafish microscopy. Its documented inference workflow takes microscopy data through pretrained 3D detection, learned frame-to-frame association, ILP graph tracking, track repair and division gating, then serializes a competition submission. The repository also contains an experiment controller for managing research runs and reviews. The deployment notebook described in the README is not present in the supplied tree, so several pipeline paths point to the exact enclosing source directories available here; those implementation details are documented rather than sampled directly.
Main entry points: loop.py Experiment workflow · predict_unet_transformer.py 3D detector · biohub_tracking Cell detections · simple_node_transformer.py Transformer association
Components 14
Inputs
- Microscopy data
Detection
- 3D detector
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/scripts/predict_unet_transformer.py - Cell detections
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/src/biohub_tracking
Tracking
- Transformer association
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/src/biohub_tracking/models/simple_node_transformer.py - ILP tracking
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/src/biohub_tracking - Motion and gap repair
experiment_controller - Division gates
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/src/biohub_tracking/division_metrics.py
Submission
- Tracking graph
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/src/biohub_tracking/io.py - submission.csv
experiments/repro_003_public_0933_kaggle_slug/remote_output/tracking_repo/scripts/predict_unet_transformer.py
Research control
- Experiment workflow
experiment_controller/loop.py - Review and state
experiment_controller/review.py - Experiment records
experiment_controller/core.py
Other components
- Kaggle user
- Kaggle competition
Connections 14
- Kaggle user → Microscopy data: selects data
- Microscopy data → 3D detector: provides frames
- 3D detector → Cell detections: predicts cells
- Cell detections → Transformer association: supplies detections
- Transformer association → ILP tracking: scores links
- ILP tracking → Motion and gap repair: forms tracks
- Motion and gap repair → Division gates: repairs tracks
- Division gates → Tracking graph: gates divisions
- Tracking graph → submission.csv: serializes graph
- submission.csv → Kaggle competition: submits results
- Kaggle user → Experiment workflow: manages runs
- Experiment workflow → Experiment records: updates state
- Experiment workflow → Review and state: requests review
- Review and state → Experiment records: records verdict