Towards General Purpose Object Detection: Deep Dense Grid Based Object Detection

Abstract

Object detection is one of the most challenging and very important branch of computer vision. Some of the challenging aspect of a detection network is the fact that an object can appear anywhere in the image, be partially occluded by another object, might appear in crowd or have greatly varying scales. Consequently, we propose a fine grained and equally spaced dense grid cells throughout an input image be responsible of detecting an object. We re-purpose an already existing deep state-of-the-art detector or classifier into deep and dense detector. Our dense object detector uses binary class encoding and hence suitable for very large multi-class object detector. We also propose a more flexible and robust non-max suppression implementation to filter out redundant detection of same object. As a result of our dense object detection implementation we have managed to increase YOLOv2's performance on Pascal VOC 2007 and COCO datasets by +2.3% and +7.2% mean average precision (mAP) respectively.