FUTURE · FINAL REGISTRY · ORGANIZER RULING APPLIED

Final Model Candidate Registry and Execution Queue

The registry preserves accepted, current, next, planned, parked, deferred, and excluded candidates while CPU preparation runs ahead of strictly serial RTX4090 work.

PROJECT MAP50 AUTHORITYCPU PREP PARALLELGPU EXECUTION SERIAL
Accepted baselines · ranked by project mAP50 only
RankModelProject mAP50Execution status
1Co-DINO ViT-L + O3650.716519509ACCEPTED CANONICAL BASELINE
2DEIM-D-FINE-X + O3650.563005103ACCEPTED BASELINE
3Corrected DINO R500.561574311ACCEPTED BASELINE
4Stable-DINO R500.540012673ACCEPTED BASELINE
5RT-DETR R500.416151944HISTORICAL ACCEPTED BASELINE
Current, next, and planned candidates
CandidateComplianceExecution state
Stable-DINO Swin-L 384 · 4-scaleLICENSE CLEAREDCURRENT · FULL TRAINING IN PROGRESS
RT-DETRv4-X + training-only DINOv3 teacherCONDITIONAL — OFFLINE TEACHER ONLYNEXT · TEACHER ARTIFACT READY · WAITING FOR GPU HANDOFF
MI-DETR Swin-Large-384LICENSE CLEAREDCPU PREP CANDIDATE
D-FINE-X + Objects365LICENSE CLEAREDCPU PREP CANDIDATE
Parked and deferred candidates
CandidateStatusReason
LW-DETR XLargePARKED — EXECUTION/HARNESS ISSUERecovery limit reached after CUDA compatibility work.
RF-DETR-LPARKED — EXECUTION/HARNESS ISSUEPrediction bridge / evidence-integrity issue.
Mr. DETR Align* Swin-LPARKED — EXECUTION/HARNESS ISSUESource compatibility chain / recovery limit.
Relation-DETR FocalNet-LPARKED — RTX4090 CAPACITYGenuine RTX4090 BS1 capacity OOM.
Stable-DINO FocalNet-HugeDEFERREDApproximately 689M parameters; high RTX4090 capacity risk and poor opportunity-cost profile. Revisit only if Stable-DINO scaling evidence justifies it.
Permanently excluded candidates
Candidate classStatusBasis
RF-DETR XLEXCLUDED — PML 1.0Custom non-OSI PML 1.0 license.
RF-DETR 2XLEXCLUDED — PML 1.0Custom non-OSI PML 1.0 license.
Any final detector using DINOv3 as backboneEXCLUDED — DINOv3 BACKBONEOrganizer prohibits DINOv3 as final detector backbone.
Any pretrained model/data under custom restrictive licensingEXCLUDED — NON-OSIOrganizer requires OSI-approved licensing.
Continuous baseline queue
StageExecution
CURRENTStable-DINO Swin-L canonical training; CPU preparation of future candidates.
NEXTRT-DETRv4-X teacher equivalence → RTX4090 qualification → 58-epoch canonical training.
PLANNEDMI-DETR Swin-Large; D-FINE-X; other justified viable candidates.
DEFERREDStable-DINO FocalNet-Huge only if later Stable-DINO scaling evidence justifies the cost.
POST-BASELINEFull shared audit of ALL canonical models → Pareto analysis → separate Tuning, Ensemble, and Internal-MoE pools.
Retained historical / variant registry entries
CandidateStateBoundary
DEIMv2 · OSI student + offline DINOv3 teacherVARIANT AUDIT REQUIREDPotentially permitted; exact recipe provenance required before GPU work.
DEIMv2 · DINOv3 final backboneEXCLUDED — DINOv3 BACKBONEOrganizer prohibition applies.
DINO 5-scale Swin-L + Objects365HISTORICAL RESERVERetained for decision history; no active GPU allocation.
RT-DETR R101-O365 controlHISTORICAL RESERVERetained for decision history; no active GPU allocation.
RT-DETRv2 reserveHISTORICAL RESERVERetained for decision history; no active GPU allocation.