Most current contactless palm vein (PV) recognition systems are designed for indoor scenes. However, in the wild, the performance of traditional region of interest (ROI) extraction methods can be significantly degraded or even fail. To tackle the challenges above, in this paper, we propose a method for extracting ROI specifically designed for PV in the wild. A top-down keypoint classification network (KCNet) for PV ROI extraction is designed. To verify the validity of our model, we annotate five keypoint datasets of SCUT_PV_v1, CASIA, TJ_PV and VERA, and for the first time, construct a PV dataset for wild scenes called SCUT_PV_Wild. Extensive experiments demonstrate that our method can achieve stable and efficient ROI extraction and get remarkable results on five annotated datasets.