154 — Spot the Difference by Object Detection

Wu et al (1801.01051v1)

Read on 21 January 2018
#computer-vision  #object-detection  #CNN  #books 

If you have two images and you run an object detection algorithm on both, you are likely to identify scenes as “identical” even when they’re not: The classic “Jaws” artwork would return shark, woman, and ocean just like a photograph of an oceanographer holding a baby nurse shark would. Classic object detection systems are not helpful when differentiating similar pictures.

In contrast, using a CNN, the algorithm presented in this paper reliably spots the differences between very slight variations of the same image. This is useful, the authors show, for detecting the differences between digital and printed copies of a book cover; detecting slight differences between versions of a printed book, etc.

This system not only detects subtle differences; it also gives fine-grained information about the location and nature of the change.

This sort of technology is very useful for cases in which automated book listings on e-commerce sites mis-cite the version of locality of a book: By training the network on one version of the book, it’s possible to notice even very slight changes to the cover of the book, common in the cases of localization or versioned updates.