Abstract  Weakly-supervised object detection (WOD) is a  challenging problems in computer vision. The key  problem is to simultaneously infer the exact object  locations in the training images and train the object  detectors, given only the training images with weak  image-level labels. Intuitively, by simulating the  selective attention mechanism of human visual  system, saliency detection technique can select  attractive objects in scenes and thus is a potential  way to provide useful priors for WOD. However,  the way to adopt saliency detection in WOD is not  trivial since the detected saliency region might be  possibly highly ambiguous in complex cases. To  this end, this paper first comprehensively analyzes  the challenges in applying saliency detection to  WOD. Then, we make one of the earliest efforts to  bridge saliency detection to WOD via the self-paced  curriculum learning, which can guide the learning  procedure to gradually achieve faithful knowledge  of multi-class objects from easy to hard. The  experimental results demonstrate that the proposed  approach can successfully bridge saliency detection  and WOD tasks and achieve the state-of-the-art  object detection results under the weak supervision