DCAM: Disturbed Class Activation Maps for Weakly Supervised Semantic Segmentation

• DOI: 10.2139/ssrn.4310669
• Publication Date: 2023-01-01T10:34:33Z
• Authors: Jie Lei; Guoyu Yang; Shuaiwei Wang; Zunlei Feng; Ronghua Liang

In the field of weakly supervised Semantic segmentation (WSSS), Class Activation Maps (CAM) are typically adopted to generate pseudo masks. Yet, we find that crux unsatisfactory masks is incomplete CAM. Specifically, as convolutional neural networks tend be dominated by specific regions in high-confidence channels feature maps during prediction, extracted CAM contains only parts object. To address this issue, propose Disturbed (DCAM), a simple yet effective method for WSSS. We adopt binary cross-entropy (BCE) loss train multi-label classification model. Then, disturb some map and retrain model order enhance encoder addition, softmax (SCE) branch employed increase attention target classes. Once converged, extract DCAM same way The evaluation on both PASCAL VOC MS COCO shows not generates high-quality (6.2% 1.4% higher than benchmark models), but also enables more accurate activation object regions. code available at https://github.com/gyyang23/DCAM.