What makes Paris look like Paris? (2012)

Wow--didn't expect my old paper to re-surface in 2018.

Besides being a neat visualization, one remarkable thing (to me) is that this is one of the few computer vision tasks where the state-of-the-art is still the old hand-designed descriptors--specifically HOG. The attempts to re-do this work with deep learning haven't worked so well. The closest I know is https://arxiv.org/abs/1506.06343 , which I've never even seen applied to cities. The problem is actually that deep nets have too much invariance: they can classify whether a facade is in Paris or not, but there's no easy way to separate out different kinds of facades.

If anyone has questions about this work, post below and I'll try to answer.

Anecdotally, after having lived in Paris for ~6 months, one of the most noticeable aspects of the city is its relative uniformity. Unlike Berlin, London, or Prague, you can wander for awhile in Paris and essentially see the same style of architecture and urban landscape repeated. It has something of an “endless” effect on your time perception.

There is a way to test the validity of their result. Go to figure 6 of the paper, in which they have a set of "Extracted Visual Elements" for each city. Cover the captions under each set, then see if you can pick the city. If their results are valid, shouldn't the essence of the city be in their selected photos, allowing a person to identify the city by looking at them?

I picked London, but none of the others (which might just reflect my lack of knowledge of each city). What do others see in figure 6?

I've always been amazed at how at the epicenter of a large city or event there are many normal looking things that contribute subtly or inadvertently to the overall awe. It's not like everything in Times Square is magical, yet the sum of them is.

I wonder how many of these more ordinary objects or store fronts you could remove and still keep the aesthetic wonder of the place. Take out the garbage cans, the street lights, the cross walks, bring the buildings a little closer together or father apart. At what point does it's quality simply change?

Can this type of software be used the other way around, to turn any picture of some building into a picture of that building in Paris?

That's an effect I've seen with many deep learning things but that I don't understand.

I'm not sure Paris does look like Paris. It's quite famous for being a let down: https://en.m.wikipedia.org/wiki/Paris_syndrome

Although that publication says 2015, this work goes back to at least 2012: https://news.ycombinator.com/item?id=4374542.

I think the smell of Paris is very unique

a lot of black folks?