Marine DebrisCompetition research · v7

ROADMAP · 24 AUG 2026 · TAIWAN TIME (UTC+8)

Pareto Completion & Final Submission Roadmap

The project has moved from broad detector exploration into strong model completion → ensemble optimization → crop rescoring → final submission system. Finalist work remains conditional on selection on 3 September.

MI-DETR V2 ACTIVETEAM RESULT ≈0.7701REPRODUCTION HANDOFFPRELIM DEADLINE 01 SEPFINALIST CONDITIONAL
Competition timeline · Taiwan time (UTC+8)
MilestoneDateStatus
Preliminary submission deadline2026-09-01 23:59:59DEADLINE
Announcement of finalists2026-09-03GATE
Finalist submission deadline2026-09-15 23:59:59CONDITIONAL DEADLINE
COMPLETED

Strong model controls

Co-DINO1280 and RT-DETRv4-X1280 are accepted. Co-DINO1280 demonstrated a local and hidden scale-transfer gain.

ACTIVE

MI-DETR Effective-Scale AUG V2

Finish the 12-epoch AWS L4 run before making a causal replacement decision.

REPRODUCTION

Team system handoff

Reconstruct the approximately 0.7701 hidden pipeline before treating it as fully canonical local authority.

Current decision tree · Preliminary phase

1 · Finish MI-DETR1280 Effective-Scale AUG V2Record the complete 12-epoch trajectory and compare against MI-DETR1024 = 0.7154639
2 · Complete teammate artifact handoffReproduce the approximately 0.7701 team system with exact artifacts, preprocessing, scoring, and submission provenance
3 · Retire Co-DINO Full-Aug V1Preserve failure evidence; do not continue the scale-cancelling composed V1
4 · Run Co-DINO Effective-Scale AUG V2Fresh Objects365 initialization; no final scale-normalizing resize; correct partial accumulation window
5 · If MI-DETR V2 underperformsRun full MI-DETR1280 scale-control; do not auto-start a 24-epoch V2 retrain
6 · Freeze strongest Co-DINO + MI-DETR variantsChoose by canonical validation, hidden evidence, and contribution to the reproduced final system
7 · Test replacements inside rescorer pipelineMeasured substitutions only; retain RT-DETRv4 if it stays additive
8 · Internal freezeTarget: 31 August
9 · Final preliminary submission · DEADLINE2026-09-01 23:59:59 Taiwan time
Immediate variant decisions
Variant / branchCurrent stateNext sufficient action
MI-DETR1280 Effective-Scale AUG V2ACTIVE · best 0.708067 at epoch 6finish epochs 10–12
Co-DINO1280 Full-Aug V1FAILED · SUPERSEDEDretire; do not continue
Co-DINO1280 Effective-Scale AUG V2PLANNED / NEXTfresh controlled run after bounded GPU qualification
MI-DETR1280 scale-controlINCOMPLETE / DEPRIORITIZEDrun full control only if V2 remains clearly below 1024
RT-DETRv4-X1280COMPLETE · ACCEPTEDretain only if additive in rescorer pipeline
Broad detector sweepNOT CURRENT PRIORITYavoid blind additions
strong model completion → ensemble optimization → crop rescoring → final submission system
Crop-rescorer and team handoff status
ItemCurrent authority
ComponentConvNeXt-Tiny crop rescorer · PROVEN HIGH-VALUE COMPETITION COMPONENT
Observed evidenceapproximately 0.7701 AIdea hidden · team pipeline · reproduction handoff in progress
Fully controlled fallbackCo-DINO1280 + MI-DETR1024 WBF · local 0.7468375071854423 · hidden 0.7587924
Handoff contentscheckpoints · validation/preliminary predictions · exact ensemble recipe · crop preprocessing · score formula · final submission SHA · documentation
Artifact conventionsmall provenance in Git; large prediction/checkpoint artifacts in LFS/shared storage; datasets externally

IF SELECTED AS FINALIST · starts after 3 September

Finalist announcement · GATE2026-09-03 · this entire branch happens only if selected
Stronger crop-rescorer refinementRevalidate against the frozen finalist detector and fusion distribution
Leakage-safe calibration / simple stackingOnly if a safe OOF or frozen fit/selection strategy exists and measured gain justifies complexity
Measured class/area gating or tiny specialistRetain only if evidence supports a net final-system benefit
Additional controlled Effective-Scale experimentsOnly if a specific unresolved hypothesis remains
Cross-team ensemble ablationCompare shared artifacts; do not assume components transfer without measurement
Final submission · CONDITIONAL DEADLINE2026-09-15 23:59:59 Taiwan time
Meta Stacker V1 · finalist-phase guardrails
DecisionAuthority
Conceptual inputCo-DINO + MI-DETR + optional RT-DETRv4 → WBF → crop rescorer → Meta Stacker
Possible featuresclass ID · area fraction · aspect ratio · detector confidences · WBF score · supporter count · IoU/agreement · crop score · optional density
Outputconfidence / ranking only · no class reassignment · no box geometry modification
Initial complexitylogistic regression or small gradient-boosted tree; shallow MLP only if justified
Leakage-safe fittingOOF detector predictions if practical, or a carefully frozen fit/selection partition
Stop ruleskip if safe fitting is not cheap; retain only for material frozen-validation gain over WBF + crop rescorer