Performances in two tables in README are different?

[laisimiao]

Table1: Table2:

And in table2, NL+BBOX performance is lower than BBOX, what's the discovery?

[568xiaoma]

Table 1 shows the performance reported in the paper. Table 2 shows the performance of the released checkpoint.

We use mixed modal data (NL, BBOX, NL+BBOX) to train the model. Empirically, the more NL+BBOX samples, the better NL+BBOX tracking performance. Thus, NL+BBOX reference does not absolutely perform better than BBOX reference for our tracker. In our paper, we adopt a relatively balanced sample proportion to achieve good performance across different modal references.

[laisimiao]

So no checkpoint corresponding to performance in the paper release? I think the community prefers the consistent checkpoint to plot figures in their future research.

[568xiaoma]

Thank you for your attention and advice. The raw results are consistent with the paper, which provides the reference of our tracker for their future research. But, the original checkpoints were broken during migration, so we retrain our model and release the last checkpoint.

[laisimiao]

No wonder. Thank you and your nice work.