Figure 3. The proposed end-to-end framework for WSSS. We use a Transformer backbone as the encoder to extract feature maps. The
initial pseudo labels are generated with CAM [59] and then refined with the proposed PAR. In the AFA module, we derive the semantic
affinity from MHSA in Transformer blocks. AFA is supervised with the pseudo affinity labels derived from the refined label. Next, we
employ the learned affinity to revise the pseudo labels via random walk propagation [2, 1]. The propagated labels are finally refined with
PAR as the pseudo labels for the segmentation branch.