Joint Optimization of Object Region Localization and Recognition/Detection Model Learning

点击翻译按钮获取中文摘要

We addressed a method to locate target objects and to learn object recognition and detection models simultaneously from given large collection of images with labels of target objects. The method does not require bounding rectangles or object regions of training images but just object labels, which are easily prepared by manual annotation, or easily obtained from Flickr or closed captioned videos. Object region localization functionality enables image synthesis and more precise object recognition and detection