This made me think of something, so bear with me for a second. I took Andrew Ng's Machine Learning class on Coursera, and one of the things he talked about was how, given a certain data set and a logistic regression approach, the algorithm might decide to just always give it a value of 1, since 98% of the time the value is one.

The examples given here were all predicted to have 95%+ similarity, but it seems to me that given this dataset, I could be reasonably convinced that it just assumes all images are reasonably similar. In other words, the author provides no counter-examples that demonstrate that the algorithm can detect dissimilar images, such as a photo of a car and a clipart of pizza. I would expect that to have very low similarity, and given the algorithm's demonstrated ability to discern between similar and dissimilar images, I would be more convinced of perceptual image hashing as a viable technique.