FlyingThings3 is a synthetic dataset for optical flow, disparity, and scene flow evaluation, consisting mainly of everyday objects flying along 3D trajectories, and contains 25,000 stereo frames generated from ground truth data.

The dataset is not focused on a specific task (like KITTI) or limited to natural scenes (like Sintel), but relies on randomness and a large number of rendered assets to generate orders of magnitude more data than any existing options, without the risk of duplication or saturation.