In view of its important application value, background modeling is studied so widely that many techniques have emerged, which mainly concentrate on the selections of the basic model, the granularity of processing, the components in a framework, etc. However, the quality of samples (QoS) for training has long been ignored. There are two aspects regarding this issue, which are how many samples are suitable and which samples are reliable. To tackle the “how many” problem, in this paper, we propose a convergent method, coined Bi-Variance (BV), to decide an appropriate endpoint in the training sequence. In this way, samples in the range from the first frame to the endpoint can be used for model establishment, rather than using all the samples. With respect to the “which” problem, we construct a pixel histogram for each pixel and subtract one from each bin (called number of intensity values (NoIV-1)), which can efficiently get rid of outliers. Furthermore, our work is plug-and-play in nature, so that it could be applied to diverse sample-based background subtraction methods. In experiments, we integrate our scheme into several state-of-the-art methods, and the results show that the performance of these methods in three indicators, recall, precision, and F-measure, improved from 4.95% to 16.47%, from 5.39% to 26.54%, and from 12.46% to 20.46%, respectively.