Visual comparison of Imagenet pre-trained features of ResNet-34 and EfficientNet-b0. (a) Shows a reference image. Each row contains 25 patches with minimum distance from a reference patch. Rows of patches are sorted in order of semantically most similar (row b) to least similar (row l). Row (b) corresponds to patches extracted from EfficientNetb0-block-2. Similarly, row (c) to (l) corresponds to features extracted from ResNet34-block-1, EfficientNet-b0-block-1, ResNet34-block-2, EfficientNet-b0-block-4, EfficientNet-b0-block-3, ResNet34-block-3, ResNet34-block-4, and block 5 to 7 of EfficientNet-b0, respectively.