Rank 19 (silver medal) solution of the Kaggle featured competition TrackML Particle Tracking Challenge
This repository only contains the code for predicting a single event. See CFlow for scaling the code to all 125 testing events.
- Cluster hits using DBSCAN on transformed data using various parameters to generate a pool of track candidates. This step is in
/mymodule/cluster2.py - Remove duplicate tracks from the pool of candidates. Create a track object for each candidate. Perform helix fitting and outlier removal. This step is in
/mymodule/track.py - Merge tracks. This step is in
/mymodule/merger.py - Above procedures are repeated several times with various parameters, focusing on clustering tracks with longer length first.
The full pipeline is in pipeline.ipynb