bad performance of Attention RPN in base training

[open-mmlab]

bad performance of Attention RPN in base training

[linyq17]

In my view, attention rpn requires high-quality (large and full objects are preferred) support instances for training and needs to carefully set the fine-tuning parameters, since attention rpn tends to overfit under few shot settings.
The mmfewshot implementation of attention rpn only supports 2 ways. Since the sampling logic of attention rpn is query aware that means the support samples are selected according to the query images in the mini-batch, while methods like meta-rcnn the support samples are selected independently of query images in the mini-batch. In attention rpn, the query images in the mini-batch can not be used as support images, so 10 shot instances only can select 9 shot for support instance and 1 shot for query.

[leader402]

I have the same problem. Can you provide normal training parameters?

Environment is single 3090