ECT: Enhancement Cross Training for Weakly Supervised Semantic Segmentation

In recent years, Weakly Supervised Semantic Segmentation (WSSS) has garnered significant attention, as it enables pixel-level segmentation using only image-level labels. However, current WSSS methods typically rely on extracting Class Activation Map (CAM) from a classification network as the initial localization cues, which are often narrow and fragmented. In this paper, we demonstrate that more regions can be activated by roughly applying enhancement functions on CAM. Specifically, we propose an Enhancement Cross Training (ECT) approach for WSSS, which involves non-learning enhancement functions and a Cross Training process for integrating enhancement functions into the learnable CAM generation network. By cross-training two identical CAM generation models, ECT allows CAM to expand with its own localization information. Experiment on PASCAL VOC 2012 shows that our method is competitive with existing state-of-the-art methods.