1) This is a very narrow case of the general image search problem. To quote the paper, "The categories of queries addressed, products, lends itself well to the type of local feature detectors that we employed to generate the underlying graph." It is, at least, unclear how well the approach performs for other domains.

2) Product image search seems to call, in general, for returning a single result. They don't report how often the first result is correct. The closest stat they report improving the precision of the top 3 results from 2.19/3 (using a baseline of Google Image Search) to 2.8/3.

3) The baseline of Google Image Search seems unreasonably naive, since precision would probably be better if the search were limited to (or favored) known shopping sites with consistent structure (like Amazon or Wikipedia).

In summary, the paper's approach is technically interesting might have more general applicability. But I'll only believe it when I see the evidence, and I find their experiment unconvincing for the reasons enumerated above.

DanielTunkelang

Owen Thomas

Munjal

MunjalShah