Bounding Part Scores for Rapid Detection with Deformable Part Models

Computing part scores is the main computational bottleneck in object detection with Deformable Part Models. In this work we introduce an efficient method to obtain bounds on part scores, which we then integrate with deformable model detection. As in [1] we rapidly approximate the inner product between a weight vector and HOG-based features by quantizing the HOG cells onto a codebook and replace their inner product with the lookup of a precomputed score. The novelty in our work consists in combining this lookup-based estimate with the codebook quantization error so as to construct probabilistic bounds to the exact inner product. In particular we use Chebyshev’s inequality to obtain probably correct bounds for the inner product at each image location. We integrate these bounds with both the Dual-Tree Branch-and-Bound work of [2,3] and the Cascade-DPMs of [4]; in both cases the bounds are used in a first phase to conservatively construct a shortlist of locations, for which the exact inner products are subsequently evaluated. We quantitatively evaluate our method and demonstrate that it allows for approximately a twofold speedup over both [2] and [4] with negligible loss in accuracy.