Question about how to design dataset

[Simon-Stma]

I would like to ask again, my dataset has more small targets (remote sensing dataset), and the size difference between different objects is large (and after referring to the VOC dataset, I found that the target in the original map is a larger proportion, which is different from my dataset), I choose Novel class to contain smaller objects and larger objects (more robust). But in the process of meta-training, no matter how I adjust the parameters, there is still a normal map of Base class and almost 0 map of Novel class, maybe it is not the problem of parameters, but my support size is 320x320 by default, should I reduce this value to better support meta-training? What is a better suggestion to solve this problem?

[GuangxingHan]

Thanks very much for the follow-up works. That's a very interesting and challenging problem.

As we know Faster R-CNN has two stages, can you examine the results of proposal generation for novel classes? In this way, we can identify whether the problem comes from localization or classification.