Estimating heights from photo collections: a data-driven approach

A photo can potentially reveal a tremendous amount of information about an individual, including the individual's height, weight, gender, ethnicity, hair color, skin condition, interests, and wealth. A {\em photo collection} -- a set of inter-related photos including photos of many people appearing in two or more photos -- could potentially reveal a more vivid picture of the individuals in the collection. In this paper we consider the problem of estimating the heights of all the users in a photo collection, such as a collection of photos from a social network. The main ideas in our methodology are (i) for each individual photo, estimate the height differences among the people standing in the photo, (ii) from the photo collection, create a people graph, and combine this graph with the height difference estimates from the individual photos to generate height difference estimates among all the people in the collection, (iii) then use these height difference estimates, as well as an a priori distribution, to estimate the heights of all the people in the photo collection. Because many people will appear in multiple photos across the collection, height-difference estimates can be chained together, potentially reducing the errors in the estimates. To this end, we formulate a Maximum Likelihood Estimation (MLE) problem, which we show can be easily solved as a quadratic programming problem. Intuitively, this data-driven approach will improve as the number of photos and people in the collection increases. We apply the technique to estimating the heights of over 400 movie stars in the IMDb database and of about 30 graduate students.