Backdrop Purging is a common pre-processing step in computer vision and video processing for object tracking, people recognition, and other tasks. Several successful background-subtraction algorithms have recently been proposed, however nearly all of the best-performing ones are supervised. The availability of some annotated frames of the test video during training is critical to their performance. As a result, there is no literature on their performance on completely "unseen" videos. We provide a new supervised background-subtraction technique for unseen films (BSUV-Net) based on a fully-convolutional neural network in this paper. The current frame and two background frames collected at various time scales, along with their semantic segmentation maps, are fed into our network. We also offer a new data-augmentation strategy that mitigates the influence of illumination differences between the background frames and the current frame in order to limit the risk of overfitting. In terms of F-measure, recall, and precision, BSUV-Net beats state-of-the-art algorithms assessed on unseen videos on the CDNet-2014 dataset.