[D] What's SOTA in efficient one-shot detection for a single reference image?

To clarify, I don't mean the usual in-the-wild object detection task that's like "here's a thousand random photos, put bounding boxes around all the bicycles". I mean a much smaller-scale task that's more like "here's a thousand images that passed through a company fax machine, show me which ones have that company's logo somewhere on the page". As in, you have a clean reference copy of the exact image you're looking for within the larger images, but it might be …

bicycles detection image images machine machinelearning mean object photos random reference scale show sota through