Great work. The paper is well written.
This method seems like a natural candidate for object detection, since it comes with the natural capability to focus on a subarea of an image and compare that in isolation to a reference image. I wonder if you or somebody else is already working on that?
Another direction that interests me is to use this method to select "high-confidence samples" and with them the original network could be further improved (comparable to this paper Few-shot Object Detection https://arxiv.org/pdf/1706.08249.pdf). If I understand it correctly the RNN controller is not altered at all by the one-shot examples.
Great work. The paper is well written.
This method seems like a natural candidate for object detection, since it comes with the natural capability to focus on a subarea of an image and compare that in isolation to a reference image. I wonder if you or somebody else is already working on that?
Another direction that interests me is to use this method to select "high-confidence samples" and with them the original network could be further improved (comparable to this paper Few-shot Object Detection https://arxiv.org/pdf/1706.08249.pdf). If I understand it correctly the RNN controller is not altered at all by the one-shot examples.