Boosting Deformable Part Model by Sample Sharing and Outlier Ablation

The deformable part model (DPM) achieves the best performance on some well known datasets in terms of object detection. Literature springs up to study the success of such a model and hence various methods are proposed to improve it. Yet one import issue, the sensitivity to outliers of the hinge loss, has not been fully studied. In this paper, we take two initiatives to handle this problem: 1) we propose to share samples of one component to others by similarity; 2) we give samples different weights according to their costs. The model is better trained with our proposed method, and we boost the performance of the newly released voc-release 5 [6] model on the challenging VOC 2007 dataset.