Class-aware Overview-first Look-closely-next Model for Single-Stage Weakly Supervised Semantic Segmentation

Weakly Supervised Semantic Segmentation (WSSS) aims to generate pixel-wise predictions using only image-level labels. However, conventional pipelines still suffer from inherent limitations: Class Activation Maps (CAMs) only highlight the most discriminative regions of objects rather than full extent, multi-stage refinement methods demand extensive computational resources, and conventional single-stage models relying on plain or pyramid architectures inherently struggle to expand Effective Receptive Fields (ERFs) and lack explicit top-down semantic guidance. To tackle these challenges, we propose Class-aware Overview-first Look-closely-next (CaLo), a single-stage WSSS framework that fundamentally adapts a top-down architecture for WSSS by explicitly injecting semantic guidance into local feature refinement. Inspired by OverLoCK's overview-focus paradigm, CaLo adopts an SR-MCT-style Vision Transformer as Base-Net to produce mid-level tokens and multi-class tokens, and applies a Deep Stage Decomposition with two lightweight branches: Overview-Net that rapidly yields a coarse context prior, and Focus-Net that progressively refines features. The core of our novel top-down design is Class-aware Gated Aggregator (CAGA) that realizes context-mixing dynamic kernels, where multi-class tokens compute class-to-feature affinities and dynamically gate token mixing, thereby expanding the ERFs and providing explicit top-down guidance to recover object extents without heavy computation. Trained from scratch on PASCAL VOC 2012 and MS COCO 2014 using only image-level labels, CaLo delivers strong accuracy and efficiency, surpassing the prior state-of-the-art single-stage methods. Our results demonstrate that coupling overview-driven global priors with explicit class-aware dynamic token aggregation is a simple yet powerful recipe for scalable single-stage WSSS.