mpi_inf_3dhp

我们提出了一种基于 CNN 的方法，用于从单个 RGB 图像进行 3D 人体姿态估计，该方法解决了仅在极其有限的公开可用 3D 姿态数据上训练的模型的泛化性有限的问题。仅使用现有的 3D 姿势数据和 2D 姿势数据，我们通过学习特征的转移在已建立的基准上展示了最先进的性能，同时也推广到了野外场景。我们进一步介绍了一个新的训练集，用于从真实人类的单眼图像中进行人体姿态估计，该训练集具有使用多摄像机无标记运动捕捉系统捕获的基本事实。它补充了现有的语料库，在姿势、人类外观、服装、遮挡和视点方面具有更大的多样性，并能够扩大增强范围。我们还提供了一个涵盖室外和室内场景的新基准，并证明我们的 3D 姿势数据集显示出比现有注释数据更好的野外性能，结合 2D 姿势数据的迁移学习进一步改进。总而言之，我们认为将表示的迁移学习与算法和数据贡献结合使用对于一般 3D 身体姿势估计至关重要。

We propose a CNN-based method for 3D human pose estimation from a single RGB image, which addresses the limited generalization problem of models trained on only extremely limited publicly available 3D pose data. By solely utilizing existing 3D and 2D pose data, we demonstrate state-of-the-art performance on established benchmarks via feature transfer learning, while also enabling generalization to in-the-wild scenarios. We further introduce a novel training set for human pose estimation from monocular images of real humans, which has ground truth annotations captured using multi-camera markerless motion capture systems. This dataset complements existing corpora, exhibiting greater diversity in poses, human appearance, clothing, occlusions and viewpoints, and allows for expanded augmentation ranges. We also present a novel benchmark that covers both outdoor and indoor scenes, and demonstrate that our 3D pose dataset achieves better in-the-wild performance than existing annotated datasets, with further improvements via transfer learning integrated with 2D pose data. In summary, we argue that combining transfer learning of representations with algorithmic and data contributions is critical for general 3D human pose estimation.

搜集汇总

背景与挑战

背景概述

mpi_inf_3dhp是一个用于3D人体姿态估计的数据集，专注于从单个RGB图像进行姿态预测，旨在解决现有3D数据有限导致的模型泛化性问题。该数据集通过多摄像机无标记运动捕捉系统捕获，提供了丰富的姿势、外观和场景多样性，包括室内外环境，并支持迁移学习以提升野外性能。