Records · kv:2610.00045v1

Reproduction of IDEA-Research/GroundingDINO: zero-shot COCO detection with GroundingDINO-T

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-24

Independent reproduction · Artifact evaluation · Computer vision

PaperReproduction of: original code

Abstract

Ran demo/test_ap_on_coco.py at the pinned commit with the GroundingDINO_SwinT_OGC config and the released groundingdino_swint_ogc.pth weights on all 5000 COCO val2017 images, zero-shot: the 80 category names joined into one text prompt, top 300 boxes per image, box AP from pycocotools. Weight and annotation SHA-256 hashes are printed in the run; bert-base-uncased was cached beforehand and the run was offline. One change to the code: the multi-scale deformable attention CUDA extension (ms_deform_attn_cuda.cu) used value.type() in two AT_DISPATCH macros, which current PyTorch no longer accepts, so it now uses value.scalar_type(); the extension was compiled for this GPU with CUDA 12.8. PyTorch 2.11, transformers 4.49. One NVIDIA RTX 5060 Ti.

Sealed report

Cite this record

@misc{kveritas261000045v1,
  title={Reproduction of IDEA-Research/GroundingDINO: zero-shot COCO detection with GroundingDINO-T},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00045v1},
  url={https://kveritas.org/records/2610.00045v1},
}

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.