PWC-Net is a compact and effective CNN model for optical flow estimation, designed based on established principles such as pyramidal processing, warping, and the use of a cost volume. The model uses learnable feature pyramids to warp the second image's features towards the first image, constructing a partial cost volume that is processed by a CNN to estimate optical flow. PWC-Net outperforms existing methods on the MPI Sintel final pass and KITTI 2015 benchmarks, achieving lower average endpoint errors (EPE) and running at about 35 frames per second (fps) on Sintel resolution images. It is 17 times smaller in size and easier to train than FlowNet2, and provides better accuracy and speed compared to SpyNet. The model's effectiveness is demonstrated through ablation studies and visual results, showing its ability to handle large motions, severe occlusions, and strong shadows.PWC-Net is a compact and effective CNN model for optical flow estimation, designed based on established principles such as pyramidal processing, warping, and the use of a cost volume. The model uses learnable feature pyramids to warp the second image's features towards the first image, constructing a partial cost volume that is processed by a CNN to estimate optical flow. PWC-Net outperforms existing methods on the MPI Sintel final pass and KITTI 2015 benchmarks, achieving lower average endpoint errors (EPE) and running at about 35 frames per second (fps) on Sintel resolution images. It is 17 times smaller in size and easier to train than FlowNet2, and provides better accuracy and speed compared to SpyNet. The model's effectiveness is demonstrated through ablation studies and visual results, showing its ability to handle large motions, severe occlusions, and strong shadows.