Detecting Advertising Materials via Multi-Scale Instance Segmentation Network

In this paper, we introduce a completely new application in computer vision: detecting advertising materials in commodity pictures. Different from the conventional detection tasks, the success of advertising materials detection greatly relies on successfully capturing the semantic and contour information of candidate targets. In our work, we firstly adopt a fully convolutional instance segmentation network to capture the semantic information and link information of pixels. Secondly, ASPP module and multi-scale prediction structure are introduced to handle materials with various scales. Then, we jointly optimize the network with semantic loss, link loss, and contour loss, to obtain finer detection results. Finally, we provide a dataset specially targeting at advertising materials detection. Experiments on this dataset show the promising performance of the proposed method.