Figure 2: Overview of ReTax-Seg. The upper panel summarizes the training pipeline. The visual
backbone and forward model are learnable. CLIP and SAM remain frozen. Boundary quality scores
the segmentation outputs. The lower panel details the multiscale visual backbone, the frozen Mobile-
SAM geometry branch, representative semantic atoms, and the frozen CLIP text-prior construction.
The symbol θ∗ denotes the weights associated with dataset d. The star marks the selected trained
d
checkpoint. For example, θ∗ represents the Cityscapes-specific weights. Snowflakes and flames
City
indicate frozen and learnable components, respectively.