- TL;DR: Spatial information at last layers is not necessary for a good classification accuracy.
- Abstract: Intuitively, image classification should profit from using spatial information. Recent work, however, suggests that this might be overrated in standard CNNs. In this paper, we are pushing the envelope and aim to further investigate the reliance on and necessity of spatial information. We propose and analyze three methods, namely Shuffle Conv, GAP+FC and 1x1 Conv, that destroy spatial information during both training and testing phases. We extensively evaluate these methods on several object recognition datasets (CIFAR100, Small-ImageNet, ImageNet) with a wide range of CNN architectures (VGG16, ResNet50, ResNet152, MobileNet, SqueezeNet). Interestingly, we consistently observe that spatial information can be completely deleted from a significant number of layers with no or only small performance drops.