The real world and the virtual world are generally considered to be two separate worlds. Leotta and Mundy, however, have changed that. In their paper, they present an approach to track vehicles in surveillance videos based on a virtual computer-aided design (CAD) model of the vehicle. Instead of using statistically derived features from real-world input data, the model being used comes from a man-made virtual CAD mesh.

The paper describes the approach in detail, and very visually, and then presents an evaluation on a small dataset, followed by some discussion. The idea is very novel and the results are impressive. However, a major concern is the benchmarking on a self-made small dataset. I recommend reading the paper anyway, as it’s really easy to read and the approach is clearly outlined.