few-shot

[Javier]

Thanks for the amazing work. I think one-shot & few-shot detection are going to change the game of object detection.

Problem Statement:

So far we can add multiple labels of one super class to the one-shot detector (for example, we could add multiple cars and bikes, so we know that classes 0-10 are cars and 11-20 are bikes).

Questions:

1. Is there any way to add few images to one class and somehow averaging the "Feature matching and alignment" of the results to get a more robust prediction?
2. Any way to combine the weights of multiple features before applying the head classifier?
3. Or maybe that's not even necessary, since we can group the labels (i.e., 0-10 --> car)?

I would appreciate any hint on that and would try to help in the implementation as well if that can be of any help.

[Anton Osokin]

Hi Javier, thanks for the kind words!
The main focus of our work was specifically on the one-shot setting. We only dealt with few-shot only by merging the detection boxes coming from several class representatives before NMS.
It indeed might be possible to do better, in particular, we tried to align representatives from the same class to each other and aggregate features, but the detection quality gains were always at most marginal compared to the increased runtime cost.

Best,
Anton