Enhancing Weakly Supervised Semantic Segmentation via Deep and Shallow Spatial Clues

Weakly supervised semantic segmentation (WSSS) leverages less informative supervision, such as image-level annotations, to reduce reliance on costly pixel-level labels. Existing methods often focus on salient features or rely on complex multistage processes to bridge the gap to fully supervised approaches. Nevertheless, traditional WSSS methods, while circumventing the need for pixel-level annotations, frequently underutilize spatial information. To address this limitation, we propose a novel framework incorporating a set of spatialized losses that explicitly enhance the use of both deep and shallow spatial cues, bringing performance closer to fully supervised methods. Unlike conventional approaches, our method avoids global average pooling in loss computation, preserving spatial details. We introduce a pixel-wise binarization loss to reinforce the distinction between foreground and background, improving model generalization. Additionally, we integrate the Segment Anything Model (SAM) to generate class-agnostic masks, which are then used to enforce spatial consistency in model predictions during training. Furthermore, our spatialized classification loss is specifically designed to leverage image-level labels while maintaining spatial integrity. These innovations collectively lead to substantial improvements in pseudo-label precision and segmentation accuracy without introducing complex architectural modifications. Experiments on the PASCAL VOC 2012 dataset demonstrate that our method outperforms existing CNN-based WSSS approaches, showcasing the effectiveness of our spatialized loss functions.