4 ms·
A Google Prototype for a Precision Image Search
- gojomo 19y agoLooking at the source paper, "PageRank for Product Image Search", it's clear this is the same approach Peter Norvig described at SS08. Namely, start with a large set of candidate images based on textual context. Then, create a similarity graph between them using extracted features that are somewhat robust against scaling and other transformations. Then, find the central clusters of that graph, picking the 'most representative' images in the clusters as good, unique results.
- gojomo 19y agoPS #1: If Munjal Shah of Riya is quoted accurately -- “I think what they’re trying to accomplish is largely impossible... Our belief is, there is not large-scale solutions.” -- then Google is lucky to face such easily-discouraged competitors.
- pixcavator 19y agoTo be fair, they tried first THEN got discouraged. I also suspect that they are talking about different things.
- gojomo 19y agoPS #2: Because TechCrunch included a DocStoc embedding of the source WWWConf paper in its story, the DocStoc page is right now the #2 result at Google for the paper title, "PageRank for Product Image Search". Scribd, fire up your SEO and MA/KA engines!
- pixcavator 19y agoIt not about improvement of image search in general (aka visual image search, CBIR etc), but about improvement of Google image search.