Records · kv:2610.00046v1

Reproduction of WongKinYiu/yolov9: COCO val2017 evaluation of YOLOv9-S, M, C and E

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-24

Independent reproduction · Artifact evaluation · Computer vision

PaperReproduction of: original code

Abstract

Evaluated the four released converted checkpoints (yolov9-s, -m, -c and -e from release v0.1) with the repository's val.py at the pinned commit on COCO val2017 (5000 images, labels from the yolov5 coco2017labels-segments archive), using the README's settings: image size 640, confidence 0.001, IoU 0.7, --save-json, with box mAP computed by pycocotools. Batch size 32, except 16 for YOLOv9-E, which does not fit in 16 GB of GPU memory at 32. Weight file SHA-256 hashes are printed in the run. The figure/ folder is withheld as hash-only entries. Changes to the environment only: PyTorch 2.11 (CUDA 12.8) with TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 to load the pickled checkpoints, NumPy 2.3 (np.trapz) and Pillow 9.5 (FreeTypeFont.getsize), plus IPython which models/common.py imports. One NVIDIA RTX 5060 Ti.

Sealed report

Cite this record

@misc{kveritas261000046v1,
  title={Reproduction of WongKinYiu/yolov9: COCO val2017 evaluation of YOLOv9-S, M, C and E},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00046v1},
  url={https://kveritas.org/records/2610.00046v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.