We show that the way inference is performed in few-shot segmentation tasks\nhas a substantial effect on performances -- an aspect often overlooked in the\nliterature in favor of the meta-learning paradigm. We introduce a transductive\ninference for a given query image, leveraging the statistics of its unlabeled\npixels, by optimizing a new loss containing three complementary terms: i) the\ncross-entropy on the labeled support pixels; ii) the Shannon entropy of the\nposteriors on the unlabeled query-image pixels; and iii) a global KL-divergence\nregularizer based on the proportion of the predicted foreground. As our\ninference uses a simple linear classifier of the extracted features, its\ncomputational load is comparable to inductive inference and can be used on top\nof any base training. Foregoing episodic training and using only standard\ncross-entropy training on the base classes, our inference yields competitive\nperformances on standard benchmarks in the 1-shot scenarios. As the number of\navailable shots increases, the gap in performances widens: on PASCAL-5i, our\nmethod brings about 5% and 6% improvements over the state-of-the-art, in the 5-\nand 10-shot scenarios, respectively. Furthermore, we introduce a new setting\nthat includes domain shifts, where the base and novel classes are drawn from\ndifferent datasets. Our method achieves the best performances in this more\nrealistic setting. Our code is freely available online:\nhttps://github.com/mboudiaf/RePRI-for-Few-Shot-Segmentation.\n