Marine Debris Roadmap v5

CURRENT · STATE 11 AUGUST 2026 · FULL TRAINING IN PROGRESS

Stable-DINO Swin-L Training + RT-DETRv4-X Handoff

Stable-DINO Swin-L is the only active GPU experiment. RT-DETRv4-X is prepared as the immediate next target and cannot begin GPU work before Stable-DINO reaches final PASS.

IN PROGRESSPROJECT APPROXIMATE BATCH-16RTX4090 GPU WORK SERIAL
0.716519509strongest accepted project mAP50
0.628511427Stable-DINO latest validated mAP50
Epoch 5 / 12latest validated point
RT-DETRv4-Xnext GPU target
Accepted-baseline leaderboard · project mAP50 only
RankAccepted modelProject mAP50
1Co-DINO ViT-L + O3650.716519509
2DEIM-D-FINE-X + O3650.563005103
3Corrected DINO R500.561574311
4Stable-DINO R500.540012673
5RT-DETR R500.416151944
STRONGEST ACCEPTED

Co-DINO ViT-L + O365

Accepted canonical baseline at project mAP50 0.716519508806631. It remains the current reference, not a guaranteed final winner.

CURRENT · TRAINING

Stable-DINO Swin-L 384 · 4-scale

Fixed 1024 × 1024 canonical run on RTX4090. Latest validated display epoch: 5 at project mAP50 0.628511427.

NEXT · READY

RT-DETRv4-X

Training-only DINOv3 teacher artifact acquired. Waiting for Stable-DINO final PASS and independent handoff verification.

Validated raw display-epoch trajectories
Display epochStable-DINO Swin-L mAP50Co-DINO mAP50
10.4362036330.518863306
20.5352598210.596071553
30.5787153210.649201281
40.6172631260.661208320
50.6285114270.668345779
Stable-DINO Swin-L canonical runtime recipe
FieldCurrent value
VariantStable-DINO Swin-L 384 · 4-scale
Resolutionfixed 1024 × 1024
Parameters218,352,190
GPURTX4090
Batchphysical 2 · gradient accumulation 8 · nominal 16 images/update
PrecisionFP32 parameters · FP16 autocast · GradScaler enabled
EMAdisabled
ProvenancePROJECT APPROXIMATE BATCH-16
Exact physical-BS16 equivalenceNO

RT-DETRv4-X handoff and canonical execution

Stable-DINO final PASS
Independent handoff verification
DINOv3 teacher compatibility / conversion
HF-vs-native numerical equivalence
Frozen teacher wrapper verification
RTX4090 BS1 full-topology smoke
Physical-batch calibration
100 successful optimizer-update qualification
Checkpoint / resume verification
Batch-policy resolution
58-epoch canonical RT-DETRv4 training
RT-DETRv4-X prepared state
FieldVerified / planned value
StudentHGNetv2-B5 / X-tier · 62,916,630 parameters
TeacherFrozen training-only DINOv3 ViT-B/16 · LVD-1689M
HF repositoryfacebook/dinov3-vitb16-pretrain-lvd1689m
HF revision5931719e67bbdb9737e363e781fb0c67687896bc
Artifactmodel.safetensors · 342,662,192 bytes
SHA2569a21ac3df0c63839d62612dda6f454d816c25611cc7a52966ed5a5a94921dc8b
Canonical target58 epochs · 1024 × 1024 · 34 classes
Checkpoint selectionproject map_50 · preserve last, best project-map50, and final
StatusTEACHER ARTIFACT ACQUIRED · WAITING FOR STABLE-DINO FINAL PASS
CPU / NETWORK / DISK

Prepare candidate N+1

While candidate N trains, verify and pin source, audit license, acquire and hash checkpoints, create environments, validate dependencies, prepare configs/adapters, and run CPU/static tests.

RTX4090 · STRICTLY SERIAL

Train candidate N

GPU preflight → BS1 smoke → physical-batch calibration → numerical qualification → checkpoint/resume → canonical training → standalone verification → next candidate.

Reproducibility strategy

  • frozen dataset split
  • pinned upstream revisions
  • checkpoint SHA256
  • qualified per-model environments
  • deterministic dataset adapters where required
  • fixed seed

Acceptance evidence

  • standalone project metric evaluation
  • best reproducible project mAP50
  • evidence freeze after accepted runs