SAL-Net: Self-Supervised Attribute Learning for Object Recognition and Segmentation

Figure 2

The architecture of the proposed SAL-Net. It contains (1) a shared feature encoding module (the green block), in which features from multiple CNN layers are encoded as attribute features under the supervision of attribute relationships; (2) a segmentation module (the orange block), which maps the attribute features into coarse-to-fine segments; (3) a classification branch (the blue block), which predicts category labels with the combination of category and attribute description.