Putting the Anchors Efficiently: Geometric Constrained Pedestrian Detection

Anchor box mechanism is of vital importance for deep network based object detection. Current object detectors put anchors uniformly in the entire image, so the false positives along with training and testing costs are increased significantly. In this work, we break through the homogeneity limitation in anchor putting by introducing geometric constraints for pedestrian detection. We first deduce the relationship between the height of the pedestrian and its location in image plane using geometrical priors, for which we only need to know the rough relative height between camera and pedestrian. As a result, we narrow down the distribution space of anchors with a certain pixel height from 2D to 1D. In implementation, we propose a novel Geometric Constrained Loss, by which the new anchor mechanism is embedded into the deep learning architecture. To further remove false positives in inference, Geometric Constrained Suppression is introduced. Complemented with two effective prediction modules, Dilated Double Shots Detector is designed to achieve better performance. Comprehensive experiments conducted on KITTI and Caltech datasets validate the superior performance and generalization ability of our method for both one-stage and two-stage methods, in both training and inference phases.