Large-Scale Particle Image Velocimetry (LSPIV) has been successfully applied for stream flow and flood measurements. In most LSPIV studies, the water surface has been approximated as a two-dimensional planar surface. Errors may arise due to this approximation for flows in steep channels and with complex riverbed topography, especially for small mountain rivers. In this study, a Stereo Imaging based LSPIV system (SI-LSPIV) is developed to reconstruct the three-dimensional topography and water surface distribution, in addition to surface velocity measurements. The SI-LSPIV system consists of two synchronized Raspberry Pi camera modules that can capture stereo images autonomously. Uncalibrated rectification and point cloud methods are implemented for surface reconstruction. Detailed three-dimensional surface topography as well as water level distribution can be obtained, which facilitates image ortho-rectification and wetted-area delineation for LSPIV calculation. Over the past years, an unattended SI-LSPIV system has been deployed in Longxi River, Sichuan, China. One major flow event was captured and the results were presented to demonstrate that LSPIV analysis aided with three-dimensional water surface distribution can improve velocity measurement and stream discharge estimation for torrential mountain river flows.