Marine DebrisCompetition research · v8

16 SEP 2026

MODELS · PARETO CANDIDATES · LOCAL CANONICAL VS HIDDEN

Standalone Pareto Board

Local canonical mAP50 and AIdea preliminary score are separate evidence columns. A dash means no hidden result is claimed or tracked here.

14 RANKED ENTRIESLOCAL MAP50AIDEA WHEN VERIFIEDLW-DETR EXCLUDED

14ranked standalone entries

0.725874859best standalone local mAP50

0.7373246best standalone AIdea

3completed 1280 controls

34object classes

Standalone Pareto board · ranked by canonical local mAP50
Rank	Variant	Resolution / augmentation	Local mAP50	AIdea hidden	Status / why it matters
1	MI-DETR Swin-L · 1280 Effective-Scale AUG V2 E13	1280 · Effective-Scale V2	0.7258748594	0.7373246	BEST · ACCEPTED · HIDDEN VERIFIED · E13 selected
2	Co-DINO ViT-L + O365 · 1280 scale-control	1280 · flip-only	0.7216258164	0.7312659	COMPLETE · ACCEPTED · HIDDEN VERIFIED
3	RT-DETRv4-X · 1280	1280 · scale-control	0.7166061138	—	ACCEPTED LOCAL STANDALONE · HIDDEN PENDING · tiny/crowd recall
4	Co-DINO ViT-L + O365 · 1024	1024 · accepted cohort	0.7165195088	—	COMPLETE · ACCEPTED 1024 COHORT
5	MI-DETR Swin-L · 1024	1024 · RandomFlip-only	0.7154639176	0.7310095	COMPLETE · ACCEPTED · HIDDEN VERIFIED
6	Mr. DETR Align Swin-L + O365	accepted standalone	0.6902207126	—	COMPLETE · accepted
7	Stable-DINO Swin-L	accepted standalone	0.6710156116	—	COMPLETE · accepted
8	D-FINE-X + O365	accepted standalone	0.6245201439	—	PARETO · strong recall / FP trade-off
9	RF-DETR-L	accepted standalone	0.6149111868	—	PARETO · accepted historical standalone
10	DEIMv2-L + DINOv3	accepted standalone	0.5870749950	—	PARETO · lower-cost trade-off
11	DEIM-D-FINE-X + O365	accepted standalone	0.5630051027	—	COMPLETE · accepted
12	Corrected DINO R50	R50 · 4-scale	0.5615743110	—	DIAGNOSTIC · low-resource role
13	Stable-DINO R50	R50 · 4-scale	0.5400126733	—	COMPLETE · accepted
14	RT-DETR R50	R50	0.4161519439	—	DIAGNOSTIC · inference-speed role

Accepted 1024 cohort · preserved local comparison set
Rank	Variant	Canonical local mAP50	Evidence boundary
1	Co-DINO ViT-L + Objects365	0.716519508806631	ACCEPTED 1024 COHORT
2	MI-DETR Swin-L	0.7154639176139567	ACCEPTED COHORT · fresh local reproduction 0.7154703347623489
3	RT-DETRv4-X	0.7043832595246645	ACCEPTED 1024 COHORT
4	Mr. DETR Align Swin-L + O365	0.6902207126154145	ACCEPTED 1024 COHORT
5	Stable-DINO Swin-L	0.6710156116233708	ACCEPTED 1024 COHORT
6	D-FINE-X + O365	0.6245201439298089	ACCEPTED 1024 COHORT
7	RF-DETR-L	0.6149111867950395	ACCEPTED 1024 COHORT
8	DEIMv2-L + DINOv3	0.5870749950004589	ACCEPTED 1024 COHORT
9	DEIM-D-FINE-X + O365	0.5630051027049761	ACCEPTED 1024 COHORT
10	Corrected DINO R50 4-scale	0.5615743110418727	ACCEPTED 1024 COHORT
11	Stable-DINO R50	0.5400126732683338	ACCEPTED 1024 COHORT
12	RT-DETR R50	0.4161519439428504	ACCEPTED 1024 COHORT

Scale-up standalones and pending candidate
Variant	Canonical local mAP50	AIdea preliminary / hidden	Status / interpretation
MI-DETR Swin-L 1280 Effective-Scale AUG V2 E13	0.7258748594176209	0.7373246	ACCEPTED STANDALONE · completed
RT-DETRv4-X 1280	0.7166061137871418	not tracked	ACCEPTED LOCAL STANDALONE · HIDDEN PENDING · +0.0122228542624773 vs 1024
MI-DETR1024 strict-3/3-ignore + synthetic-883 U812	0.714006930956	not available	CANDIDATE PENDING HIDDEN · LOCAL REGRESSION · NOT PROMOTED

Resolution scaling evidence · Co-DINO
Variant	Local canonical mAP50	AIdea hidden	Interpretation
Co-DINO 1024	0.716519509	0.7183351	accepted historical control
Co-DINO 1280 scale-control	0.721625816	0.7312659	+0.0051 local / approximately +0.01293 hidden

Resolution scaling evidence · MI-DETR
Variant	Local canonical mAP50	AIdea hidden	Interpretation
MI-DETR 1024 scale-control	0.7154639176	0.7310095	accepted control
MI-DETR 1280 Effective-Scale V2 E13	0.7258748594	0.7373246	+0.010411 local / +0.006315 AIdea