The issue of referring results inconsitency

[liuyang-ict]

Thanks for your very impressive work for OVD. However, there is an inconsistent issue in the REC results of MM-GroundingDINO between Table 5 and the original paper (e.g. 66.0 top-1 precision of refcoco-val in Table 5, but 53.* top-1 precision in the original paper of Table 5). Are there any typos here？

[fushh]

Thanks for your interest.
As elaborated in Section C of Appendix, we have improved the inference pipeline of REC tasks due to task discrepancy. Thus, the performance in our Table 5 is higher than the one in the original paper.

[liuyang-ict]

wow, that's a nice idea. May I inquire if it is possible to share the refined validation files of the REC?

[fushh]

please see mmdet/models/detectors/grounding_dino.py

[liuyang-ict]

I see. Thanks for your timely reply~~