I have trained a FRCNN on my custom dataset and I just want the output of RPN and ROI-pool.
I have used resnet50 FPN as the backbone.
For every image, I just want the region proposals from the network and their corresponding feature maps so that I can run my custom classifier that I have developed separately on this features that am getting.
The issue is -
initially I used a simple RESNET - non-FPN architecture and I was able to do what I wanted since the output of the backbone was a single set of feature maps, however when using FPN am getting multiple feature maps the name and shape of which am showing below :
0 torch.Size([1, 256, 200, 200])
1 torch.Size([1, 256, 100, 100])
2 torch.Size([1, 256, 50, 50])
3 torch.Size([1, 256, 25, 25])
pool torch.Size([1, 256, 13, 13]) ‘’’
Now I don’t understand which of these feature maps to use and I saw a formula k=k0+log2(sqrt(wh)/224) but am not sure how and where to use this.
Also when I trained using simple Non-FPN architecture i had used the same backbone for my FRCNN as well as the custom classifier so now if I were to re-train my classifier using ResnetFPN that yields multiple feature maps how should I do that?
ps : the custom classifier is an OCC-one class classifier