Traffic forecasting has been an active research field in recent decades, and with the development of deep-learning technologies, researchers are trying to utilize deep learning to achieve tremendous improvements in traffic forecasting, as it has been seen in other research areas, such as speech recognition and image classification. In this study, we summarize recent works in which deep-learning methods were applied for geospatial data-based traffic forecasting problems. Based on the insights from previous works, we further propose a deep-learning framework, which transforms geospatial data to images, and then utilizes the state-of-the-art deep-learning methodologies such as Convolutional Neural Network (CNN) and residual networks. To demonstrate the simplicity and effectiveness of our framework, we present a formulation of the New York taxi pick-up/drop-off forecasting problem, and show that our framework significantly outperforms traditional methods, including Historical Average (HA) and AutoRegressive Integrated Moving Average (ARIMA).