Mixed High-Order Attention Network for Weakly- Supervised Semantic Segmentation

Abstract

Existing weakly-supervised semantic segmentation methods using image-level annotation typically rely on Class Activation Map (CAM) to locate the object regions. However, the previous works concentrate only on the calculation of first-order attention and ignored the exploration of high-order attention on semantic segmentation tasks, resulting in the response maps generated by the classification network is only focusing on discriminative object parts. In this paper, we use high-order statistics mechanism to capture the fine-grained information in the picture, so that the classification network can also locate the detailed part of the object. We propose to add the mixed high-order attention module to the classification network to further enrich the attention information. Extensive experiments on PASCAL VOC 2012 dataset demonstrate this mechanism has excellent performance on weakly-supervised semantic segmentation.