Figure 3.

The graphical representation of our CRF. ${\mathcal{F}}_c$ and ${\mathcal{F}}_f$ represent the feature embedding functions for the coarse and the fine segmentation networks. Our model consists of the scene-level unary term ${\mathsf{\Phi }}_c(\mathbf{x};\mathbf{I})$, the object-level unary term ${\mathsf{\Phi }}_f(\mathbf{x};\mathbf{I})$, and the pixel-level pairwise term $\mathsf{\Psi }(\mathbf{x};\mathbf{I},\mathbf{D})$.