Identity initialization sets a layer’s weights so it begins as a near-no-op: information passes forward unchanged, and training learns useful deviations from that mapping. section 3.5
Dilated-convolution context modules
In a segmentation context module, layers initially pass information through unchanged. Training then learns contextual refinements, stabilizing the module without explicit skip connections or architectural changes. section 3.5
