Mitigating Undisciplined Over-Smoothing in Transformer for Weakly Supervised Semantic Segmentation

A surge of interest has emerged in weakly supervised semantic segmentation due to its remarkable efficiency in recent years. Existing approaches based on transformers mainly focus on exploring the affinity matrix to boost CAMs with global relationships. While in this work, we first perform a scrupulous examination towards the impact of successive affinity matrices and discover that they possess an inclination toward sparsification as the network approaches convergence, hence disclosing a manifestation of over-smoothing. Besides, it has been observed that enhanced attention maps tend to evince a substantial amount of extraneous background noise in deeper layers. Drawing upon this, we posit a daring conjecture that the undisciplined over-smoothing phenomenon introduces a noteworthy quantity of semantically irrelevant background noise, causing performance degradation. To alleviate this issue, we propose a novel perspective that highlights the objects of interest by investigating the regions of the trait, thereby fostering an extensive comprehension of the successive affinity matrix. Consequently, we suggest an adaptive re-activation mechanism (AReAM) that alleviates the issue of incomplete attention within the object and the unbounded background noise. AReAM accomplishes this by supervising high-level attention with shallow affinity matrices, yielding promising results. Exhaustive experiments conducted on the commonly used dataset manifest that segmentation results can be greatly improved through our proposed AReAM, which imposes restrictions on each affinity matrix in deep layers to make it attentive to semantic regions.