The thesis "Beyond Pixels: Exploring New Representations and Applications for Motion Analysis" by Ce Liu, submitted to the Department of Electrical Engineering and Computer Science at MIT, focuses on advancing motion analysis beyond the pixel level and exploring new applications. The author addresses the limitations of traditional motion analysis, which often fails to accurately model pixel grouping and lacks ground-truth data. In the first part of the thesis, Liu introduces a human-assisted motion annotation system that uses layers as an interface for users to interact with video sequences, providing ground-truth motion data. This system also enables the detection and magnification of small motions, making them more visible to human observers. Additionally, the thesis explores contour presentation for analyzing motion in textureless objects under occlusion, demonstrating that simultaneous boundary grouping and motion analysis can overcome challenges faced by traditional pixel-wise motion analysis. The second part of the thesis focuses on matching local image structures rather than intensity values. Liu proposes SIFT flow, a method that establishes dense and semantically meaningful correspondences between images across scenes by matching pixel-wise SIFT features. Using SIFT flow, a new framework for image parsing is developed, transferring metadata such as annotation, motion, and depth from a large database to an unknown query image. This framework is applied to predict motion from a single image and synthesize motion via object transfer. The thesis also introduces a nonparametric scene parsing system using label transfer, which outperforms state-of-the-art techniques based on training classifiers. Overall, the thesis contributes to the field of motion analysis by advancing representation methods and exploring novel applications, enhancing the ability of computers to understand and interpret motion in digital images.The thesis "Beyond Pixels: Exploring New Representations and Applications for Motion Analysis" by Ce Liu, submitted to the Department of Electrical Engineering and Computer Science at MIT, focuses on advancing motion analysis beyond the pixel level and exploring new applications. The author addresses the limitations of traditional motion analysis, which often fails to accurately model pixel grouping and lacks ground-truth data. In the first part of the thesis, Liu introduces a human-assisted motion annotation system that uses layers as an interface for users to interact with video sequences, providing ground-truth motion data. This system also enables the detection and magnification of small motions, making them more visible to human observers. Additionally, the thesis explores contour presentation for analyzing motion in textureless objects under occlusion, demonstrating that simultaneous boundary grouping and motion analysis can overcome challenges faced by traditional pixel-wise motion analysis. The second part of the thesis focuses on matching local image structures rather than intensity values. Liu proposes SIFT flow, a method that establishes dense and semantically meaningful correspondences between images across scenes by matching pixel-wise SIFT features. Using SIFT flow, a new framework for image parsing is developed, transferring metadata such as annotation, motion, and depth from a large database to an unknown query image. This framework is applied to predict motion from a single image and synthesize motion via object transfer. The thesis also introduces a nonparametric scene parsing system using label transfer, which outperforms state-of-the-art techniques based on training classifiers. Overall, the thesis contributes to the field of motion analysis by advancing representation methods and exploring novel applications, enhancing the ability of computers to understand and interpret motion in digital images.