This paper introduces a novel method called *multi-scale orderless pooling* (MOP-CNN) to improve the robustness of deep convolutional neural network (CNN) activations for recognition tasks. MOP-CNN extracts CNN activations from local patches at multiple scales, performs orderless VLAD pooling at each scale, and concatenates the results. This approach enhances geometric invariance without degrading discriminative power, outperforming global CNN activations on various datasets such as SUN397, MIT Indoor Scenes, ILSVRC2012/2013, and INRIA Holidays. The method is effective for both supervised and unsupervised tasks, achieving state-of-the-art results on challenging datasets. The paper also discusses the invariance properties of global CNN activations and compares MOP-CNN with alternative pooling methods, demonstrating its superior performance.This paper introduces a novel method called *multi-scale orderless pooling* (MOP-CNN) to improve the robustness of deep convolutional neural network (CNN) activations for recognition tasks. MOP-CNN extracts CNN activations from local patches at multiple scales, performs orderless VLAD pooling at each scale, and concatenates the results. This approach enhances geometric invariance without degrading discriminative power, outperforming global CNN activations on various datasets such as SUN397, MIT Indoor Scenes, ILSVRC2012/2013, and INRIA Holidays. The method is effective for both supervised and unsupervised tasks, achieving state-of-the-art results on challenging datasets. The paper also discusses the invariance properties of global CNN activations and compares MOP-CNN with alternative pooling methods, demonstrating its superior performance.