Two full weeks in Beijing now, and my internship is in full swing. My new “toy” at work, which I’ve been playing with every day, is TILT – Transform Invariant Low-Rank Textures, a technique for extracting 3D information out of a photo.
Here’s how it works: in photos of urban scenes, there exists an abundance of well-textured areas, otherwise known as buildings. Buildings contain lots of vertical and horizontal lines, which when expressed as pixel values, create a matrix with rows and columns that are mostly repetitive. The repetition yields a matrix of low rank (i.e. having a low rank means that there are only a few unique rows/columns in the matrix). Yet, this texture matrix is only low-rank when you look at it head-on; if viewed at an angle, the texture becomes tilted and the matrix gains a much higher rank. The key idea is that you can find the rotation that, when applied to the tilted matrix, minimizes its rank. And that rotation is what yields 3D information about objects in the photo.
Yes… chocolate ice cream served in a toilet-shaped bowl. I’ve seen everything now.
tl;dr lessons of the week:
1. Used bikes in China cost around 50RMB, or roughly a meal at McDonalds.
2. Beijing actually has a sky (not just clouds and smog). But it only appears after torrential thunderstorms.