FUTURE · CROSS-MODEL EVIDENCE GATE

Shared Model Audit + Complementarity

Compare every completed representative under one frozen evaluation protocol before selecting finalists.

PLANNED EVALUATIONDECISION GATENO COMBINED MAP/LOSS SCORE
overall mAPprecision / recallper-class APtiny / small / largecrowding / densityrare classesfalse positivesmisseslocalizationduplicate predictionserror ledgerpairwise complementarity
TRAINING HEALTH

Trajectories

Train loss, validation loss, and validation mAP across epochs.

TRAINING HEALTH

Selection landmarks

Best-mAP epoch, best-validation-loss epoch, and final epoch.

DECISION

No synthetic score

Report the trajectories and landmarks separately; do not combine mAP and loss into one score.

Select for strong individual performance plus useful nonredundancy—not headline mAP alone.