The audit separates candidate recovery from confidence ranking. It supplies the rationale for high-resolution variants, selective RT-DETRv4 use, and the now-proven-but-not-yet-fully-reconstructed crop rescorer.
MI-DETR wins 9 classes, RT-DETRv4-X 5, Mr. DETR 1, and Stable-DINO Swin-L 1.
SIZE
RT-DETRv4 leads tiny recall
RT-DETRv4-X leads recall below 0.1% image area; D-FINE-X leads the other major size bands.
DENSITY
MI-DETR leads sparse scenes
MI-DETR leads at 0–5 objects/image; Co-DINO leads the 6–10, 11–20, and >20 slices.
Shared hard-case concentration
Objects <1% · universal failures60.3%
515 of 854
Objects <1% · universal misses78.9%
315 of 399
Objects <1% · validation GT share29.6%
strong overrepresentation in residual errors
854 / 6,449universal failures
399 / 6,449universal misses
140 / 224<0.1% universal failures
124 / 224<0.1% universal misses
A · MISSING-CANDIDATE PROBLEM
Higher resolution → candidate recovery
The 12-model audit shows that universal misses are disproportionately tiny. The active 1280 experiments test whether scale can recover candidates that the current detector pool never proposes.
B · RANKING PROBLEM
Crop rescorer → ranking / precision
External teammate evidence suggests that learned visual crop features can improve the confidence ordering of candidates already proposed by detectors and fusion.
External crop-rescoring evidence · do not compare numerically with our leaderboard
External branch
val_sel
AIdea preliminary
Observed change
Clean EFS WBF
0.7497
0.7175357
reference
EFS + safe ConvNeXt-Tiny crop rescorer
0.7640
approximately 0.729
approximately +0.0143 local / +0.0115 hidden
FEATURE EVIDENCE
Visual crops were necessary
Metadata-only reranking did not provide the same gain; actual crop/image features were required.
STARTING ARCHITECTURE
ConvNeXt-Tiny
ConvNeXt-Base did not beat Tiny in the teammate branch's frozen local gate, so Tiny is the preferred V1 starting point.
COLLABORATION BOUNDARY
Transparent handoff
The teams are aligning artifacts and recipes; this does not mean their training pipelines have merged. The immediate task is reproducible reconstruction of the proven team result.
A · MISSING CANDIDATES
Resolution recovers proposals
The current 1280 controls test whether tiny objects absent from the candidate pool can be recovered.
B · RANKING / PRECISION
Fusion → crop rescoring → conditional stacking
Once candidates exist, fusion and visual crop evidence can improve ordering. A later stacker may learn trust from detector scores, agreement, fused confidence, crop evidence, geometry, and context.