Past Pixel-Box for Pedestrian Detection

Inspired by regression for detection task, We propose a new method called Pixel-Box. Prior work on regression for object detection can not match small pedestrian. Instead, Pixel-Box is only uses one feature map pixel of each channel to regress offset of boxes and then uses all feature map to predict which position has object. The model has two sibling network at the end of convolution output, but simply forward and backward once for both training and testing. It is an end to end system can directly optimized by loss function. To improve accuracy, we present a new backward method directly impact the model input and appropriate change the input image to re-forward the network. We evaluate this method on Caltech benchmark, also presenting competitive speed.