PROJECT / In progress
TraceFormer · Small-Target Detection
Complete-system validation score · 0.9541
Detecting faint moving targets in event streams with a reproducible model, postprocessing, and evaluation pipeline.
Complete system
Neural detector, LightGBM postprocessing, and trajectory rules
Independent evaluation
24 sequences, all-event scoring, and prediction-mapping audits
Results
The frozen complete system achieved a validation score of 0.954117 across 24 sequences, compared with 0.915002 for the neural detector alone in an independent evaluation on September 29, 2026.
These are local validation results. The validation set informed model and threshold selection; they do not represent official hidden-test scores or competition rankings.
Problem & method
Small targets produce sparse events amid dense background activity. Occupied-block tokens and bidirectional spatiotemporal cone attention aggregate motion neighborhoods, then map predictions back to original events. The complete system adds LightGBM postprocessing, trajectory rules, and dense-scene rescue.
My contribution
Developed the model, training, inference, and postprocessing pipeline, ran ablations and independent evaluations, and audited prediction counts, order, and event mappings.
Validation comparison
| Configuration | Validation score |
|---|---|
| Public PACT weights | 0.803205 |
| TraceFormer neural detector | 0.915002 |
| TraceFormer complete system | 0.954117 |
All use the same sequences, event mapping, and global scoring. Training and development budgets differ, so the comparison describes these particular weights and configurations. 0.954117 belongs to the frozen historical system, not every later module.
Deliverables
Runnable code, saved configurations and weights, per-sequence comparisons, and difficult-scene diagnostics. Neural-detector and complete-system metrics are reported separately.
Project gallery
