This project re-implements A-Fast-RCNN (CVPR 2017) in Caffe, using adversarial spatial dropout networks to generate hard positive training examples for object detection. The approach modifies Fast R-CNN to include an adversary network that creates occluded and deformed object appearances, forcing the detector to learn robust features. Results show a 3.2 mAP improvement over baseline Fast R-CNN on VOC 2007 using VGG16.