AGORA

具有高度真实感和高度准确性的综合数据集。在这里，我们使用4240市售的，高质量的，纹理的人体扫描在不同的姿势和自然服装; 这包括257扫描儿童。我们通过将SMPL-X身体模型 (面部和手部) 拟合到3D扫描，并考虑到服装，创建参考3D姿势和身体形状。我们通过使用基于图像的照明或渲染的3D环境渲染每张图像5到15个人，创建大约14k训练和3k测试图像，并注意使图像在物理上合理且真实。AGORA总共由173K个人作物组成。我们在此数据集上评估了用于3D人类姿势估计的现有最新方法，发现大多数方法在儿童图像上的表现不佳。因此，我们扩展了SMPL-X模型，以更好地捕捉儿童的形状。此外，我们对AGORA的方法进行了微调，并在AGORA和3DPW上显示了改进的性能，从而证实了数据集的真实性。我们在https:// agora.is.tue.mpg.de/上提供所有注册的3D参考训练数据，渲染图像和基于web的评估站点。A comprehensive dataset with high photorealism and accuracy. Here, we utilized 4240 commercially available, high-quality, textured human scans captured in diverse poses and wearing natural clothing; this collection includes 257 scans of children. We generated reference 3D poses and body shapes by fitting the SMPL-X body model (with facial and hand modules) to the 3D scans while accounting for clothing artifacts. We produced approximately 14k training and 3k test images by rendering 5 to 15 individuals per image using image-based lighting or pre-rendered 3D environments, with strict guarantees for physical plausibility and realism. AGORA contains a total of 173K individual human crops. We evaluated existing state-of-the-art 3D human pose estimation methods on this dataset, and found that most models perform poorly on images of children. Accordingly, we extended the SMPL-X model to better capture child body shape characteristics. Furthermore, we fine-tuned our proposed method on AGORA, and demonstrated improved performance on both AGORA and 3DPW, which validates the realism of this dataset. We provide all registered 3D reference training data, rendered images, and a web-based evaluation server at https://agora.is.tue.mpg.de/

搜集汇总

背景与挑战

背景概述

AGORA是一个用于3D人体姿态估计的高质量综合数据集，包含4240个市售的高质量纹理人体扫描，特别关注儿童数据。数据集通过SMPL-X模型拟合，生成了约14k训练和3k测试图像，总计173K个人作物，旨在提升3D姿态估计的准确性和真实性。