Accelerating Deformable Part Models with Branch-and-Bound

Deformable Part Models (DPMs) play a prominent role in current object recognition research, as they rigorously model the shape variability of an object category by breaking an object into parts and modelling the relative locations of the parts. Still, inference with such models requires solving a combinatorial optimization task. In this chapter, we will see how Branch-and-Bound can be used to efficiently perform inference with such models. Instead of evaluating the classifier score exhaustively for all part locations and scales, such techniques allow us to quickly focus on promising image locations. The core problem that we will address is how to compute bounds that accommodate part deformations; this allows us to apply Branch-and-Bound to our problem. When comparing to a baseline DPM implementation, we obtain exactly the same results but can perform the part combination substantially faster, yielding up to tenfold speedups for single object detection, or even higher speedups for multiple objects.