Marine Debris Roadmap v4

PAST · SHARED DIAGNOSTIC AUDIT · PASS

RT-DETR R50 vs Corrected DINO R50

A ranking, error, localization, and complementarity audit on the shared validation protocol.

FACTINFERENCENO COMPONENT-LEVEL CAUSAL CLAIM
0.5615743110418727DINO mAP50
0.4161519439428504RT-DETR mAP50
+0.14542236709902234DINO advantage
31 / 34classes won by DINO at AP50
Corrected DINO's advantage comes primarily from better confidence/ranking behavior, false-positive and duplicate suppression, broad class performance, and improved localization—not from universally higher low-threshold candidate recall.
Detection reconciliation at score threshold 0.30
MetricRT-DETRDINO
TP3,6543,444
FP7,7112,164
Precision0.3215130.614123
Recall0.5665990.534036
F10.4102390.571286
Duplicates3,287110
2,858both correct
586DINO-only correct
796RT-DETR-only correct
2,209both fail

Why RT-DETR still matters

RT-DETR retains meaningful complementary value: 796 RT-only correct GT objects, higher recall at score 0.01, and an advantage on the >20 objects/image subset point to crowded-scene specialization.

Why naive union is undesirable

The unique detections arrive with many model-specific false-positive and duplicate regions. Fusion must recover complementary objects without importing that burden.