Automated segmentation of kidney and tumor from 3D CT scans is necessary for the diagnosis, monitoring, and treatment planning of the disease. In this paper, we describe a two-stage framework for kidney and tumor segmentation based on 3D fully convolutional network (FCN). The first stage preliminarily locate the kidney and cut off the irrelevant background to reduce class imbalance and computation cost. Then the second stage precisely segment the kidney and tumor on the cropped patch. The proposed method ranks the 4th place out of 105 competitive teams in MICCAI 2019 KiTS Challenge with a Composite Dice of 90.24%.