Maximize mAP50
Select and freeze the strongest reproducible single model or justified ensemble after the shared audit, diagnosis-driven tuning, and simplicity-gated fusion.
Maximize lightweight efficiency / NetScore
Retain sufficient detection performance while optimizing the official accuracy-efficiency trade-off. The teacher comes from the strongest model or models discovered by the main track.
Performance freeze → lightweight optimization → submission
Select and freeze final performance candidate(s)Best reproducible checkpoint, configuration, inference, thresholds, and 34-class mapping.
LIGHTWEIGHT OPTIMIZATIONBegins only after the main performance candidate or candidates are frozen.
Select strongest teacherChoose from the strongest model or models discovered by the main performance track.
Select smaller student variant(s)
Knowledge Distillation
Student Pareto AuditCompare mAP vs Parameters vs GFLOPs; the smallest student is not automatically the winner.
Resolution Optimization
Precision / Quantization LadderFP32 → FP16 / BF16 → INT8 PTQ → mixed INT8 / FP16 if needed → INT8 QAT if needed.
Compute official lightweight efficiency scoreUse the competition-authoritative NetScore / efficiency formula without inventing a substitute score.
Select lightweight submission candidate
Optional pseudo-private checkOnly if enabled and still genuinely pseudo-private.
Final reproducible validation
License / compliance audit
Submission integrity + readback
Independent final review
FINAL SUBMISSION
Performance freeze
- best reproducible checkpoint(s)
- configs
- inference code
- thresholds
- 34-class mapping
- ensemble or router only if justified
Submission integrity
- correct image and category IDs
- correct bbox convention
- finite scores
- no NaN or Inf
- schema validation
- artifact readback