The paper introduces SURREAL, a new large-scale synthetic dataset for human pose, shape, and motion estimation. The dataset contains over 6 million frames generated from 3D motion capture data, with realistic images, ground truth pose, depth maps, and segmentation masks. The synthetic images are rendered using the SMPL body model and MoSh method to fit 3D motion capture data. The dataset includes diverse viewpoints, clothing, and lighting conditions, and is designed to enable accurate human depth estimation and part segmentation in real RGB images. The paper demonstrates that CNNs trained on the synthetic dataset can generalize well to real images, achieving high accuracy in both tasks. The dataset is publicly available, along with code for generating synthetic data and training models for body part segmentation and depth estimation. The paper also presents a detailed approach for generating synthetic data, including body modeling, texture, lighting, and camera parameters. The dataset is evaluated on several tasks, including human body part segmentation and depth estimation, on both synthetic and real data. The results show that the synthetic data can be used effectively for training deep networks, and that the realism of the synthetic images is sufficient to support training for complex tasks. The paper also discusses the design choices of the dataset, including the amount of data, clothing variation, and MoCap variation, and shows that the synthetic data can be used to improve performance on real data. The results demonstrate that the synthetic data can be used to train models that generalize well to real data, and that the dataset is a valuable resource for advancing person analysis using synthetic data.The paper introduces SURREAL, a new large-scale synthetic dataset for human pose, shape, and motion estimation. The dataset contains over 6 million frames generated from 3D motion capture data, with realistic images, ground truth pose, depth maps, and segmentation masks. The synthetic images are rendered using the SMPL body model and MoSh method to fit 3D motion capture data. The dataset includes diverse viewpoints, clothing, and lighting conditions, and is designed to enable accurate human depth estimation and part segmentation in real RGB images. The paper demonstrates that CNNs trained on the synthetic dataset can generalize well to real images, achieving high accuracy in both tasks. The dataset is publicly available, along with code for generating synthetic data and training models for body part segmentation and depth estimation. The paper also presents a detailed approach for generating synthetic data, including body modeling, texture, lighting, and camera parameters. The dataset is evaluated on several tasks, including human body part segmentation and depth estimation, on both synthetic and real data. The results show that the synthetic data can be used effectively for training deep networks, and that the realism of the synthetic images is sufficient to support training for complex tasks. The paper also discusses the design choices of the dataset, including the amount of data, clothing variation, and MoCap variation, and shows that the synthetic data can be used to improve performance on real data. The results demonstrate that the synthetic data can be used to train models that generalize well to real data, and that the dataset is a valuable resource for advancing person analysis using synthetic data.