Weakly Supervised Semantic Segmentation via Alternate Self-Dual Teaching
Abstract
Weakly supervised semantic segmentation (WSSS) is a challenging yet important research field in vision community. In WSSS, the key problem is to generate high-quality pseudo segmentation masks (PSMs). Existing approaches mainly depend on the discriminative object part to generate PSMs, which would inevitably miss object parts or involve surrounding image background, as the learning process is unaware of the full object structure.
In fact, both the discriminative object part and the full object structure are critical for deriving of high-quality PSMs. To fully explore these two information cues, we build a novel end-to-end learning framework, alternate self-dual teaching (ASDT), based on a dual-teacher single-student network architecture.
Methodology
The information interaction among different network branches is formulated in the form of knowledge distillation (KD). Unlike the conventional KD, the knowledge of the two teacher models would inevitably be noisy under weak supervision.
Inspired by the Pulse Width (PW) modulation, we introduce a PW wave-like selection signal to alleviate the influence of the imperfect knowledge from either teacher model on the KD process. This mechanism allows the framework to:
1. Dual-Teacher Architecture: Simultaneously leverage discriminative object parts and full object structure information through two separate teacher models.
2. Pulse Width Modulation: Dynamically select and filter knowledge from teacher models to mitigate the impact of noisy supervision signals.
3. Self-Dual Teaching: Enable mutual learning and refinement between the dual teachers and single student through alternate teaching strategies.
Experimental Results
Comprehensive experiments on the PASCAL VOC 2012 and COCO-Stuff 10K demonstrate the effectiveness of the proposed ASDT framework, and new state-of-the-art results are achieved.
The results validate that our approach successfully addresses the challenges in weakly supervised semantic segmentation by effectively combining discriminative and structural information through the novel dual-teaching mechanism.