This paper introduces a new optical flow dataset derived from the open-source 3D animated short film Sintel. The dataset addresses limitations of previous optical flow evaluation benchmarks, such as the Middlebury benchmark, by providing longer sequences, larger motions, and more complex visual effects like specular reflections, motion blur, and atmospheric effects. The dataset is publicly available and includes metrics and an evaluation website. The authors evaluate several recent optical flow algorithms and find that current top-performing methods on the Middlebury benchmark struggle with this more complex dataset, suggesting the need for further research. They also compare the image and flow statistics of Sintel to those of real films and videos, finding them to be similar. The dataset includes multiple render passes, allowing for the evaluation of flow algorithms under varying conditions. The authors also introduce perturbed sequences to detect fraud and analyze the performance of optical flow methods on different types of motion, including large motions, motion boundaries, and occluded regions. The dataset is shown to be a reasonable proxy for real-world videos, and the authors provide a detailed analysis of the statistical properties of the data. The paper also discusses the limitations of previous optical flow datasets and the importance of a centralized public comparison to encourage innovation in the field. The authors conclude that the Sintel dataset is a valuable resource for evaluating optical flow algorithms and that further research is needed to improve their performance on complex real-world scenes.This paper introduces a new optical flow dataset derived from the open-source 3D animated short film Sintel. The dataset addresses limitations of previous optical flow evaluation benchmarks, such as the Middlebury benchmark, by providing longer sequences, larger motions, and more complex visual effects like specular reflections, motion blur, and atmospheric effects. The dataset is publicly available and includes metrics and an evaluation website. The authors evaluate several recent optical flow algorithms and find that current top-performing methods on the Middlebury benchmark struggle with this more complex dataset, suggesting the need for further research. They also compare the image and flow statistics of Sintel to those of real films and videos, finding them to be similar. The dataset includes multiple render passes, allowing for the evaluation of flow algorithms under varying conditions. The authors also introduce perturbed sequences to detect fraud and analyze the performance of optical flow methods on different types of motion, including large motions, motion boundaries, and occluded regions. The dataset is shown to be a reasonable proxy for real-world videos, and the authors provide a detailed analysis of the statistical properties of the data. The paper also discusses the limitations of previous optical flow datasets and the importance of a centralized public comparison to encourage innovation in the field. The authors conclude that the Sintel dataset is a valuable resource for evaluating optical flow algorithms and that further research is needed to improve their performance on complex real-world scenes.