Evaluation on RefCOCO about pseudo-q

[jianghaojun]

The eval.sh is provided to reproduce the results reported in the paper. For RefCOCO, we report the results on val/testA/testB splits.

We need to evaluate the performance on the manually annotated val/test for fair comparison with other methods. Thus, the pseudo samples are only used in training.

You can download the manually annotated val/test from this link.

[dneirfi]

Then, unc_testA.pth, unc_testB.pth, unc_val.pth files mean the manually annotated val/test datasets?

Also, I try to reproduce the results on RefCOCO+, RefCOCOg and other datasets.
How can I get the unc+_train_imagelist_split.txt files?

[jianghaojun]

Yes, val/testA/testB are manually annotated which are original data from each dataset. The pseudo samples are only used in training.

The download link for detection results and unc+_train_imagelist_split.txt files are already provided in the README.md.

[dneirfi]

Oh, I got it.
Thanks for your quick reply!