Discriminatively Trained Part Based Model Armed with Biased Saliency

摘要

Discriminatively trained Part based Model (DPM) is one of the state-of-the-art object detectors. However, DPM complies little with real vision procedure. In this paper, we try arming DPM with biologically inspired approaches. On the one hand, we use Gabor instead of Histogram of Oriented Gradient (HOG) as low level features to simulate the receptive fields of simple cells. We show Gabor outperforms or is on par with HOG. On the other hand, we learn biased saliency of the object with the same Gabor features to simulate the search procedure of real vision. We combine DPM and biased saliency in a single Bayesian framework, which at least partially reflects the interactions between top-down and bottom-up vision procedures. We show these biologically inspired procedures can effectively improve the performance and efficiency of DPM. We present experimental results on both challenging PASCAL VOC2007 dataset and publicly available sequences.