A Classify-Before-Detect Method For Weakly Supervised Object Localization

Abstract

Classify-Before-Detect (CBD) methods provide a reliable initialization for the object detector, and are widely used in weakly supervised object localization (WSOL). This paper proposes a new CBD method for WSOL based on saliency maps. A CNN is first trained to determine the existence of objects in an image. Then, with the trained CNN model for classification, a gradient-based saliency map is obtained to generate candidate locations of objects in the image. Finally, the accurate location coordinates of objects are obtained using the special characteristics of these objects. Experiments have been conducted on the Nexar traffic light dataset, it is shown that CNNs are able to successfully classify images even though the objects occupy only a few pixels in the training images and gradient-based saliency maps provide strong resistant capability to interference. More importantly, our proposed method can locate small objects precisely, which is very difficult for the current object detectors. Moreover, our proposed method is weakly supervised and effective.