MI-DETR Swin-L 1280 Effective-Scale AUG V2 E13 · canonical local mAP50
34-CLASS OBJECT DETECTION · CANONICAL STATE
Marine Debris Image Recognition Competition
A controlled detector and ensemble program, with canonical evaluation kept separate from AIdea leaderboard evidence and diagnostic findings.
12,105 train / 25,740 boxes · 3,022 val / 6,449 boxes · zero grouped overlap
290 / 300 blinded high-confidence evaluator-FPs were debris-like
Retained in the ensemble; MI1280 replacement and recalibration were tested and rejected.
VERIFIED STATE · TAIWAN TIME (UTC+8)
Competition state is separated by authority, not collapsed into one score
0.7910702
AIdea preliminary / hidden · rank 10 / 56 · 2026-08-29
0.7258748594
MI-DETR1280 E13 canonical local · AIdea 0.7373246
892 / 1,045
PRIMARY / EXTENDED protected regions; canonical VAL remains authoritative
RECORDED STATUS
Completed decisions and last documented unresolved work
Rescorer retraining data handoff v2
883 accepted synthetic V4 images, rare-class manifests, protected exclusions, and diagnostic definitions. No later training result recorded.
MI-DETR1024 strict-3/3-ignore + synthetic-883 U812
Canonical local regression at 0.714006930956. Submission and ensemble handoff exist, but it is not promoted.
High-confidence FP / missing-annotation audit
290 / 300 debris-like; 135 / 142 low-GT-IoU cases have independent 2/3-or-3/3 same-class support.
0.7910702 exact submission lineage
29 August owner-confirmed record. Exact artifact and pipeline mapping remain unresolved in the reviewed repositories.
Evaluate the new rescorer, then test repaired MI-DETR U812 as a complementary detector
Integrate only validation-proven changes into the protected team pipeline.
Pseudo-positive training
The frozen discovery pool is not positive GT; no pseudo-positive experiment is approved or complete.
ANNOTATION-AUDIT FINDING
High-confidence evaluator false positives frequently contain visible debris.
This is strong incomplete-annotation evidence, not proof that every candidate is a missing GT box. Canonical validation remains the competition-comparison authority.
EXPLORE
Project evidence by decision surface
01Models
Accepted 1024 cohort, scale-up standalones, and candidates pending hidden evaluation.
↗02EnsemblesOwner-verified team best, reproducible fallback, and protected ensemble lineage.
↗03AnalysisMissing-annotation audit, diagnostic validation, and synthetic-data evidence.
↗04ExperimentsCompleted detector repairs, retired tuning, and a rescorer handoff with results pending.
↗05SubmissionsOwner-verified leaderboard evidence, known provenance, and pending submission candidates.
↗06RoadmapCompleted evidence, unresolved work, and the final selection boundary.
↗