About GOT-10k test set results

[researchmm]

About GOT-10k test set results

[MasterBin-IIAU]

Hi, thank you for your interest in our work.
You can try evaluating the STARK-ST50 model trained with full datasets on GOT-10K. I remember that the result is also around 0.72. Please feel free to try.

[botaoye]

Hi, thank you for your interest in our work.
You can try evaluating the STARK-ST50 model trained with full datasets on GOT-10K. I remember that the result is also around 0.72. Please feel free to try.

Thanks. I've tried STARK-ST50, and the result is fine. But I'm still confused about the reason why the STARK-S50 model trained with all dataset gains only minor improvement compared with GOT10K only but STARK-ST50 will not. Do you have any idea about this?

[MasterBin-IIAU]

Hi, I think a potential reason might be that STARK-S lacks a template update mechanism and this restricts its discriminative ability, especially when there are similar distractors. In this case, feeding model with more data shall not bring too much improvement.

[botaoye]

Hi, I think a potential reason might be that STARK-S lacks a template update mechanism and this restricts its discriminative ability, especially when there are similar distractors. In this case, feeding model with more data shall not bring too much improvement.

This may be the reason. Thanks a lot for your reply.