Marine DebrisCompetition research · v724 AUG 2026

34-CLASS OBJECT DETECTION · CANONICAL STATE

Marine Debris Image Recognition Competition

Strong model completion, ensemble optimization, crop rescoring, and a final submission system built from the fewest justified components.

PARETO CANDIDATESTEAM HANDOFF1 ACTIVE RUNPRELIM 01 SEP

BEST REPRODUCIBLE0.7587924

Co-DINO1280 + MI-DETR1024 WBF

BEST STANDALONE0.721625816

Co-DINO1280 local mAP50

ACTIVE EXPERIMENT0.708067

MI-DETR1280 V2 best so far

SHARED MISSES399

of 6,449 validation GT

COMPETITION TIMELINE · TAIWAN TIME (UTC+8)

Two submission gates, separated by finalist selection

Open full roadmap →

PRELIMINARY DEADLINE

01 SEP 2026

23:59:59 Taiwan time

FINALIST ANNOUNCEMENT

03 SEP 2026

Finalist-phase work begins only if selected.

FINALIST DEADLINE

15 SEP 2026

23:59:59 Taiwan time · conditional

CURRENT STATUS

Finish the strongest Pareto variants, then reconstruct the best team system

Open experiments →

CURRENT · ACTIVE

MI-DETR1280 Effective-Scale AUG V2

AWS L4 · epochs 10–12 active · best so far 0.708067 at epoch 6.

NEXT · PLANNED

Co-DINO1280 Effective-Scale AUG V2

Fresh Objects365 initialization after Full-Aug V1’s composed geometry failed.

PARALLEL PRIORITY

Reproduce ≈0.7701 team pipeline

Artifact, recipe, and crop-rescorer handoff are being completed; 0.7587924 remains the controlled fallback.

PROVEN COMPONENT

ConvNeXt-Tiny crop rescorer

High-value competition evidence; exact reproduction is pending, not assumed complete.

PRELIMINARY PATH

Complete variants → reconstruct system → freeze

Use measured model replacements only; retain RT-DETRv4 only if it remains additive.

IF SELECTED AS FINALIST

Refine only with evidence

Potential rescorer refinement, calibration, gating, or simple stacking—all conditional on finalist selection.

SHARED HARD-CASE DIAGNOSIS

Small objects are disproportionately represented in residual errors.

Objects below 1% image area account for 60.3% of universal failures and 78.9% of universal misses, despite representing 29.6% of validation GT. Resolution, selective recall diversity, and crop-feature ranking target different residual failures.

854universal failures

399universal misses

View the audit →

EXPLORE

Project evidence by decision surface

01Models

Standalone Pareto variants by local, hidden, resolution, and role.

02Ensembles

Reproducible WBF fallback and the stronger team rescorer result.

03Analysis

Tiny-object audit, Pareto rationale, and rescorer evidence.

04Experiments

Variant registry: active MI-DETR V2, retired V1, and next Co-DINO V2.

05Submissions

Observed team hidden evidence versus fully reproducible fallback.

06Roadmap

Completion, reconstruction, and conditional finalist decisions through 15 September.