River Surface Optical Flow Dataset for Surface Velocity Estimation

RIVER-Data Dataset DescriptionThe RIVER-Data dataset is designed to support research on image-based river surface velocity measurement and optical flow estimation. It consists of consecutive image pairs captured from natural river surfaces under a variety of flow conditions, including laminar flow, structure-induced flow, tracer-enhanced flow, and turbulent flow.Each data sample contains two temporally consecutive frames together with the corresponding optical flow field. The optical flow is represented in a two-channel format, where each pixel encodes the horizontal and vertical displacement components of motion.The dataset aims to facilitate the development, evaluation, and benchmarking of both traditional optical flow algorithms and deep learning–based motion estimation methods for hydrological applications.For each sample, the original images are provided in JPG format and follow a sequential naming convention (e.g., 0001_a.jpg and 0001_b.jpg), which together form the input image pair for optical flow estimation. The corresponding ground-truth optical flow labels are provided in the standard .flo format, storing per-pixel horizontal and vertical motion vectors and ensuring full compatibility with commonly used optical flow visualization and evaluation tools.The RIVER-Data dataset is available via Baidu Netdisk at:https://pan.baidu.com/s/143p9Vl4AY8Ji5aoYs43ZFQAccess code: vwb4To support the use of the RIVER-Data dataset, we additionally provide an optical flow estimation model, ConvFFN-Flow, which is specifically designed for motion estimation on real river surface imagery. The official implementation of this model is publicly available at:https://github.com/nxdsun/ConvFFN-Flow/