Predicting urban flow is crucial for intelligent transportation systems (ITS), but it is not easy due to several complicated elements (such as dynamic spatio-temporal dependencies, complex spatial dependence, external environment, and so on). Some studies utilize LSTM and 2D CNN networks to analyze temporal and spatial relationships independently, do not fully model spatio-temporal dependence or multiscale spatial dependence among regions. Inspired by the similarity of video analysis, we propose a new pure spatio-temporal model based on 3D convolutional neural network (3DCNN) to simultaneously capture spatio-temporal features from low-level to high-level layers, and design a grouped 3D multiscale residual strategy to directly and effectively extract multiscale spatial features. Based on these, we propose the Spatio-Temporal 3D Grouped Multiscale ResNet (ST-3DGMR), an end-to-end framework for region-based urban flow prediction. By adaptively integrating closeness and periodic spatio-temporal 3DCNN branches as well as other external factors, the ST-3DGMR can forecast future region-based inflow and outflow. To assess the performance of the proposed method, we use three representative traffic datasets. When compared to state-of-the-art techniques, experimental results show that the ST-3DGMR can lower RMSE by 2.6 %, 6.3 %, and 6.9 % on the BikeNYC, TaxiBJ, and TaxiCQ datasets, respectively.