Marine DebrisCompetition research · v7
24 AUG 2026

SUBMISSIONS · AIDEA HIDDEN · TEAM RESULT ≈0.7701

Preliminary Competition Results

Hidden scores are live competition evidence and remain distinct from local canonical mAP50. The strongest observed team result is explicitly marked reproduction-pending rather than silently treated as fully reconstructed authority.

TEAM BEST ≈0.7701REPRODUCIBLE 0.7587924LOCAL ≠ HIDDENNO RANK INFERENCE
≈0.7701best observed team hidden
0.7587924best fully reproducible hidden
0.7312659best standalone hidden
≈10reported team rank snapshot
2rejected fusion branches shown
AIdea preliminary score history
EntryTypePreliminaryStatus
Teammate pipeline: our Co-DINO1280 + MI-DETR1024 + teammate ensemble + ConvNeXt-Tiny rescorerTEAM ENSEMBLEapproximately 0.7701BEST OBSERVED · HIDDEN VERIFIED · REPRODUCTION PENDING
Co-DINO1280 + MI-DETR1024 equal-weight WBFENSEMBLE0.7587924BEST FULLY REPRODUCIBLE FALLBACK
Co-DINO ViT-L + O365 · 1280 scale-controlSTANDALONE0.7312659COMPLETE · ACCEPTED · HIDDEN VERIFIED
Co-DINO + MI-DETR class-bucket WBFENSEMBLE0.7534510NOT CANONICAL
Co-DINO + MI-DETR Cross-Model NMSENSEMBLE0.7507509SUPERSEDED
Tiny-area gated RT-DETRv4 ensembleENSEMBLE0.7484452REJECTED
MI-DETR Swin-LargeSTANDALONE0.7310095VALID
Co-DINO ViT-L + O365STANDALONE0.7183351VALID
RT-DETRv4-X · 1280STANDALONE0.7189529COMPLETE · ACCEPTED · HIDDEN VERIFIED
RT-DETRv4-X · corrected 1024STANDALONE0.7155323VALID · historical 1024
Mr. DETR Align Swin-LSTANDALONE0.7026255VALID
Stable-DINO Swin-LSTANDALONE0.6976549VALID
Competition gates · Taiwan time (UTC+8)
GateDate / deadlineAuthority
Preliminary submission2026-09-01 23:59:59ACTIVE DEADLINE
Announcement of finalists2026-09-03FINALIST PHASE STARTS ONLY IF SELECTED
Finalist submission2026-09-15 23:59:59CONDITIONAL DEADLINE

Local artifact policy

  • Evaluate every completed 1280 standalone model on canonical validation.
  • Retain standalone prediction artifacts even when no hidden submission is made.
  • Do not automatically spend a submission slot on every 1280 model.

Current preliminary-submission priority

  • 1 · Complete handoff and reproduce the approximately 0.7701 team system.
  • 2 · Test only measured replacements for the strongest Co-DINO / MI-DETR variants.
  • 3 · Retain RT-DETRv4 only if it remains additive after the rescorer pipeline is reconstructed.
  • 4 · Keep Co-DINO1280 + MI-DETR1024 WBF at 0.7587924 as the controlled fallback.
Shared artifact convention being prepared
StorageContents
Normal Gitcode · configs · tests · small reports · provenance metadata · canonical status
Git LFS / shared large-artifact mechanismvalidation/preliminary prediction JSON · .pth/.pt checkpoints
External storagefull datasets · large source archives
Per-model target bundleRUN_PROVENANCE.json · validation_predictions.json · preliminary_predictions.json · checkpoint.pth
Required provenancemodel identity · Git/upstream commit · config · checkpoint SHA256 · split SHA256 · validation score · artifact SHA256