Figure 2
From: Census-independent population estimation using representation learning
ResNet-50 ImageNet predictions and architecture. (a) ImageNet class predictions for very-high-resolution satellite image tiles. Although the classes are irrelevant, the network shows an intrinsic understanding of the difference between built-up area, vegetation, and road. (b) Diagram of ResNet-50 encoder. Residual blocks between Input and Global Average Pooling are comprised of convolutional layers with interleaved batch normalization layers.
