Person Re-Identification (ReID) across non-overlapping cameras is a\nchallenging task and, for this reason, most works in the prior art rely on\nsupervised feature learning from a labeled dataset to match the same person in\ndifferent views. However, it demands the time-consuming task of labeling the\nacquired data, prohibiting its fast deployment, specially in forensic\nscenarios. Unsupervised Domain Adaptation (UDA) emerges as a promising\nalternative, as it performs feature-learning adaptation from a model trained on\na source to a target domain without identity-label annotation. However, most\nUDA-based algorithms rely upon a complex loss function with several\nhyper-parameters, which hinders the generalization to different scenarios.\nMoreover, as UDA depends on the translation between domains, it is important to\nselect the most reliable data from the unseen domain, thus avoiding error\npropagation caused by noisy examples on the target data -- an often overlooked\nproblem. In this sense, we propose a novel UDA-based ReID method that optimizes\na simple loss function with only one hyper-parameter and that takes advantage\nof triplets of samples created by a new offline strategy based on the diversity\nof cameras within a cluster. This new strategy adapts the model and also\nregularizes it, avoiding overfitting on the target domain. We also introduce a\nnew self-ensembling strategy, in which weights from different iterations are\naggregated to create a final model combining knowledge from distinct moments of\nthe adaptation. For evaluation, we consider three well-known deep learning\narchitectures and combine them for final decision-making. The proposed method\ndoes not use person re-ranking nor any label on the target domain, and\noutperforms the state of the art, with a much simpler setup, on the Market to\nDuke, the challenging Market1501 to MSMT17, and Duke to MSMT17 adaptation\nscenarios.\n