Described is a system for automatic digital photo orientation detection.
We leverage online public photos with great content variation to extract
effective features with layout information. Classification proceeds using
an approximate nearest neighbors approach which scales well to massive
training sets, hardly compromising efficiency. We have tested the method
successfully on the largest data set to date of nearly 30,000 Flickr
photos as well as both difficult and typical consumer usage scenarios.
Though limited data are available for comparison across different
systems, the proposed system significantly outperforms a state of the art
system on a common data set.