Accelerating arrays of linear classifiers using approximate range queries

Modern object detection methods apply binary linear classifiers on Euclidean feature vectors. This paper shows that projecting feature vectors onto a hypersphere allows an approximate range query to accelerate these detectors within acceptable levels of accuracy. The expense of constructing the k-d tree used by these range queries is justified when many detectors are used. We demonstrate our acceleration technique on several existing detection systems, including a state of the art logo detector, and show that approximate range queries can detect logos at least half as well at 11× the speed of the full fidelity method.