Repository checked 16 September 2026 · Taiwan time Default branches remain at 31 August (Marine) and 30 August (Overfit). The 0.7910702 score and rank 10 / 56 are the owner-confirmed 29 August record, not a live leaderboard check. No later hidden result, finalist outcome, or active training is established by the reviewed repositories.
Completed, not queued MI-DETR1280 integration, RT-DETRv4 global inclusion, Tiny-384 trials, and scene-context ranking have recorded decisions. The separate synthetic-data rescorer handoff still has no recorded training result.
Completed integration decisions Experiment Evaluation scope Measured result Decision MI-DETR1280 E13 · blind replacement AIdea hidden · same historical pipeline 0.7689250 vs 0.7701133 · -0.0011883 REJECTED RECIPE · accepted standalone unchanged MI-DETR1280 E13 · recalibrated / dual-MI grid val_fit · 39 configurations 0.771331 vs 0.770556 · +0.000775 BELOW INTEGRATION GATE · original MI1024 retained CMY + RT-DETRv4 + Tiny/ViT reported grouped80 experiment 0.763811 vs 0.762515 · +0.001296 REJECTED GLOBAL MEMBER · no hidden result Tiny384 letterbox + ViT-S reported grouped80 experiment 0.764970 vs 0.762515 · +0.002455 BELOW INTEGRATION GATE · no hidden result LambdaRank scene context fixed w=0.7 0.775815 vs 0.776482 · -0.000666 REJECTED · unchanged ranker retained
Tiny-384 interpretation Stretch preprocessing fell from 0.777493 to 0.776497 on val_fit. Letterbox with Tiny/ViT improved the reported full split by +0.002455 but missed the gate. Resolution, geometry, data and optimization changed together, so this is not proof that 384-pixel inputs intrinsically fail.
COMPLETE · ACCEPTED MI-DETR 1280 Effective-Scale AUG V2 E13 Selected at E13 / update 9836: local 0.7258748594 and AIdea 0.7373246. E14 fell to 0.7228773224, confirming E13 as the stopping point.
RESUME PREPARED · AWS L4 Co-DINO 1280 Effective-Scale AUG V2 Portable state is verified at epoch 4 / update 6535. AWS L4 custom-op compatibility and resume preparation are complete; production training is not claimed as running.
Co-DINO ViT-L + O365 @1280 · variant registry Variant Geometry / recipe Local mAP50 Hidden Status / interpretation Scale-control flip-only · 12 epochs · BS1 × accum8 · FP16 AMP · seed 2026 0.7216258164455305 0.7312659 COMPLETE · ACCEPTED · HIDDEN VERIFIED Full-Aug V1 native-style resize/crop → final canonical 1280 resize/letterbox 0.554091 — FAILED SCIENTIFIC CANDIDATE · SUPERSEDED · shared-epoch deficit approximately -0.159641 Effective-Scale AUG V2 resize / optional crop → fixed 1280 crop/pad → no second scale-normalizing resize — — RESUME PREPARED · AWS L4 · update 6535
Co-DINO 1280 learning evidence · Scale-control vs Full-Aug V1 Epoch Scale-control Full-Aug V1 1 0.530777 0.393262 2 0.604291 0.444905 3 0.647740 0.474263 4 0.676348 0.456134 5 0.687198 0.545012 6 0.681855 0.544920 7 0.701045 0.552719 8 0.706751 0.554091 9 0.718312 0.552238
Why Co-DINO Full-Aug V1 failed V1 composed native-style stochastic resize/crop with a final canonical 1280 resize/letterbox. On resize-only paths, the final resize normalized away much of the intended scale variation. It also had a minor partial-window issue: 12,105 training images modulo accumulation8 leaves one residual microbatch that could be normalized as though it contained eight samples. The failure applies to this composed V1 implementation; it does not establish that native/full augmentation is universally bad.
MI-DETR Swin-L · variant registry Variant Geometry / precision / batch Local mAP50 Hidden Status / interpretation 1024 scale-control RandomFlip-only + centered letterbox · FP32 · BS1 × accum16 0.7154639176139567 0.7310095 COMPLETE · ACCEPTED · HIDDEN VERIFIED 1280 scale-control clean resolution control approximately 0.5875 at actual epoch 1 / optimizer update 757 — INTENTIONALLY STOPPED / DEPRIORITIZED · NOT FAILED 1280 AUG V1 resize/crop → final canonical 1280 resize — — SUPERSEDED / PRUNED · scale cancellation 1280 Effective-Scale AUG V2 E13 native-inspired resize/crop → fixed 1280 canvas · no second resize 0.7258748594 0.7373246 COMPLETE · ACCEPTED · HIDDEN VERIFIED · selected E13
MI-DETR 1280 Effective-Scale AUG V2 · completed trajectory Actual epoch Canonical mAP50 Decision 1 0.575506 interim 2 0.657886 interim 3 0.673305 interim 4 0.699389 interim 5 0.695235 interim 6 0.708067 interim 7 0.697858 interim 8 0.701916 interim 9 0.701988 interim 12 / update 9079 0.7246096288 completed 13 / update 9836 0.7258748594 BEST · ACCEPTED · AIdea 0.7373246 14 / update 10593 0.7228773224 lower than E13 · not selected 15 not run stopped after E14
MI-DETR 1280 Effective-Scale AUG V2 · accepted authority Field State Machine AWS NVIDIA L4 Precision FP16 AMP · FP32 parameters / optimizer state Batch / schedule physical 1 · accumulation16 · effective16 · selected E13 / update 9836 Validation canonical centered 1280 letterbox Reference target accepted MI-DETR1024 = 0.7154639176139567 Selection E13 peak; E14 decline; E15 not run
MI-DETR Effective-Scale V2 lesson SUCCESS. Stochastic scale augmentation is followed by a fixed 1280 crop/pad canvas with no second scale-normalizing resize. The 1024 to 1280 gain was approximately +0.010411 locally and +0.006315 on AIdea, with E13 as the peak and E14 decline confirming the selected checkpoint.
Historical crop-rescorer milestone · 0.7701133 Field Current record Component ConvNeXt-Tiny crop rescorer Local checkpoint checkpoints/safe_crop_rescorer_v1/rescorer_best.pth Checkpoint size / SHA256 111,447,199 bytes / c4af4f50ecc279b5a4e4d6672c0dab6533062064663c930ad4ba7477f1b5d5c0 Status evolution HIGH-PRIORITY EXPERIMENT → PROVEN HIGH-VALUE COMPETITION COMPONENT Evidence teammate ensemble + rescorer reached 0.7701133 AIdea hidden Caveat Historical 90,393-row submission and 90,391-row local candidate differ at a threshold boundary; later protected pipelines are documented separately.
ACCEPTED MI-DETR 1280 E13 Local 0.7258748594 and AIdea 0.7373246. The accepted 1024 comparison is 0.7154639176 / 0.7310095.
RESUME PREPARED Co-DINO Effective-Scale V2 AWS L4 resume authority is update 6535. Preparation is complete, but no running state is claimed.
RETIRED Co-DINO Full-Aug V1 Scientifically failed composed implementation; preserve evidence, but do not continue it.
Synthetic small-object and rare-class experiments Experiment Canonical result Decision SmartEraser V4 accepted scaled synthetic set 884 feasible outputs · 883 TECHNICAL_ACCEPTED · 5,231 synthetic GT · 911 intentional removals VISUAL QC PENDING; do not claim visual acceptance Mixed-883 U40 mAP50 0.7162419733 · precision 0.6600476569 · recall 0.6442859358 MODEST LOCAL GAIN / RECALL-PRECISION TRADE-OFF Rare exposure fine-tune best global U30 mAP50 0.7160629254 · +0.0005925906 NO_RARE_SIGNAL Combined small + rare tuning best mAP50 0.7148119670 COMBINED_NO_GAIN
Small-object detail Mixed-883 U40 improved recall below 0.1% from 0.084821 to 0.093750, at 0.1–<0.5% from 0.371400 to 0.397219, and at 0.5–<1% from 0.501475 to 0.517699. This is a modest local gain, not a major breakthrough.
Ignore-only detector training · completed and retired Branch Canonical evidence Verdict Strict MI+Y 2/2 ignore pilot 2,144 post-dedup ignores / 1,004 images · U40 mAP50 0.715018945739 · P 0.642758620690 · R 0.650333385021 REGRESSION Strict 3/3 ignore + synthetic-883 U812 2,156 canonical ignores / 1,010 images · 842 propagated synthetic ignores / 349 synthetic images · 812 updates REGRESSION; no second epoch authorized U812 diagnostic record canonical 0.714006930956 / 0.637606318348 / 0.650953636223 · PRIMARY 0.729297941895 / 0.735459004905 / same recall · EXTENDED 0.733167106017 recall improved; precision and both mAP views fell below baselines
Ignore-only conclusion Missing-annotation evidence and ineffective ignore-only optimization can both be true. Annotation-hole suppression alone does not explain the regression, so ignore-only detector fine-tuning is retired for now.
Rescorer retraining data handoff v2 Status Package boundary HANDOFF READY / TRAINING RESULT PENDING 883 accepted synthetic V4 images · rare-class manifests · TRAIN 3/3 negative-exclusion regions · propagated synthetic exclusions · canonical and diagnostic definitions Safety boundary information.md, no prompt.md, no pseudo-positive GT Broad propagation 18,014 exclusions · 750 synthetic images receiving exclusions · 1,337 rejected for edit overlap Intended use protect likely unlabeled debris crops from being taught as background