search

GuangxingHan / Meta-Faster-R-CNN

Code for AAAI 2022 Oral paper: 'Meta Faster R-CNN: Towards Accurate Few-Shot Object Detection with Attentive Feature Alignment'
80 stars 10 forks source link

Question about how to design dataset #17

Open Simon-Stma opened 2 years ago

Simon-Stma commented 2 years ago

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 commented 2 years ago

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.