This document proposes a framework called A-fast-RCNN that uses an adversarial network to generate hard positive examples for object detection training. It aims to address issues with occlusions, deformations, and lack of examples. The framework trains an Adversarial Sample Discovery Network (ASDN) to generate occluded and deformed examples, and an Adversarial Sample Transfer Network (ASTN) to refine localization. Evaluation on VOC2007 and VOC2012 datasets showed improved detection performance over fast R-CNN alone.