Low validation accuracy with 69M checkpoint on COCO val2017 (12% vs paper's 48.9%)

[Yohann-Ian]

Hello,
I'm trying to reproduce the results from your paper. My accuracy is low and I cannot get good results. Appreciate your help.

Questions:

1. Is the 69M_best.pt checkpoint the same model used for the paper results?
2. Are there any special inference parameters or settings required?
3. Could you share the exact validation command/configuration you used?

These are my results:
Setup:

Checkpoint: 69M_best.pt (T=1, D=4) from Google Drive
Dataset: COCO val2017 (5000 images)
Evaluation: Standard ultralytics validation protocol
Hardware: RTX 3090, CUDA 11.3

Results:

My result: mAP@0.5 = 12.0%, mAP@0.5:0.95 = 8.81%
Paper reports: mAP@0.5 = 66.2%, mAP@0.5:0.95 = 48.9%

Observations:

First ~12 classes (person, car, etc.) perform well (50-80% mAP)
Most other classes show near-zero detection
Only 4270/5000 images successfully validated

Thank you!