Person Re-Identification is a challenging task that aims to retrieve all\ninstances of a query image across a system of non-overlapping cameras. Due to\nthe various extreme changes of view, it is common that local regions that could\nbe used to match people are suppressed, which leads to a scenario where\napproaches have to evaluate the similarity of images based on less informative\nregions. In this work, we introduce the Top-DB-Net, a method based on Top\nDropBlock that pushes the network to learn to focus on the scene foreground,\nwith special emphasis on the most task-relevant regions and, at the same time,\nencodes low informative regions to provide high discriminability. The\nTop-DB-Net is composed of three streams: (i) a global stream encodes rich image\ninformation from a backbone, (ii) the Top DropBlock stream encourages the\nbackbone to encode low informative regions with high discriminative features,\nand (iii) a regularization stream helps to deal with the noise created by the\ndropping process of the second stream, when testing the first two streams are\nused. Vast experiments on three challenging datasets show the capabilities of\nour approach against state-of-the-art methods. Qualitative results demonstrate\nthat our method exhibits better activation maps focusing on reliable parts of\nthe input images.\n