Unable to reproduce the results mentioned in the paper.

[hnyang00]

Hello, I am very interested in your work and find it remarkable. So, I tried to reproduce it based on the code you provided. However, even after training multiple times with the default configuration file, I couldn't achieve the excellent results mentioned in the paper. The mAP I obtained is as follows:62.19% (D2SA),33.30% (COCOA),30.01% (KINS). This differs from the results reported in the paper.
Therefore, I have a few questions for the authors, and I hope you can help clarify my confusion:

1. Can the results mentioned in the paper be reproduced based on the configuration file given in the source code? If so, could you provide the weight files?
2. According to Is the COCOA-cls result trained from scratch or COCO pretrained? #1，I learned that you used Mask-RCNN weights trained on COCO dataset then finetuned on COCOA-cls. Could you please confirm if the same weights were used for training on the D2SA and KINS datasets, or if any other method was employed?
   Sincerely looking forward to your reply！ @trqminh