Effect of template choice on transformer

[researchmm]

Effect of template choice on transformer

[MasterBin-IIAU]

Hi, thanks for your questions. Your thinking is quite deep :)

1. "Including some context information" is inspired by the SiamFC. We actually have tried removing the background information, but the performance dropped. So we keep this design. You are right about the extreme cases where the aspect ratio is too large or too small. Maybe there is still space in the template size design.
2. Yep. The network implicitly learns to locate or "segment" the target on the template image. Besides, we also found that the learned models are quite sensitive to the "region cropping strategy" (how to crop the template and search region). If the training and testing use different strategies, the performance will drop drastically. I think this is also a potential direction for improvement.

[waterknows]

Got it. Thank you!