This paper presents a novel method for optical flow estimation called SPyNet, which combines classical spatial-pyramid approaches with deep learning. SPyNet uses a spatial pyramid structure to estimate large motions in a coarse-to-fine manner. At each pyramid level, a deep network is trained to compute the flow update, rather than minimizing an objective function. This approach results in a network that is 96% smaller and more efficient than FlowNet, with faster processing and lower memory requirements. SPyNet achieves comparable or better accuracy than FlowNet on standard benchmarks such as Sintel, KITTI, and Middlebury. The learned convolutional filters resemble classical spatio-temporal filters, providing insight into the method and potential for improvement. SPyNet is also more efficient and suitable for embedded applications. The method uses a spatial pyramid to handle large motions, with each pyramid level estimating small-motion updates. The network is trained using the same Flying Chairs dataset as FlowNet, and shows improved performance after fine-tuning. The approach combines the strengths of classical flow methods with deep learning, offering a balance between accuracy and efficiency. The results suggest that combining classical methods with deep learning can lead to more accurate and efficient optical flow estimation.This paper presents a novel method for optical flow estimation called SPyNet, which combines classical spatial-pyramid approaches with deep learning. SPyNet uses a spatial pyramid structure to estimate large motions in a coarse-to-fine manner. At each pyramid level, a deep network is trained to compute the flow update, rather than minimizing an objective function. This approach results in a network that is 96% smaller and more efficient than FlowNet, with faster processing and lower memory requirements. SPyNet achieves comparable or better accuracy than FlowNet on standard benchmarks such as Sintel, KITTI, and Middlebury. The learned convolutional filters resemble classical spatio-temporal filters, providing insight into the method and potential for improvement. SPyNet is also more efficient and suitable for embedded applications. The method uses a spatial pyramid to handle large motions, with each pyramid level estimating small-motion updates. The network is trained using the same Flying Chairs dataset as FlowNet, and shows improved performance after fine-tuning. The approach combines the strengths of classical flow methods with deep learning, offering a balance between accuracy and efficiency. The results suggest that combining classical methods with deep learning can lead to more accurate and efficient optical flow estimation.