Yoga-82

人体姿态估计是计算机视觉中用于定位关节位置的一个众所周知的问题。观察到用于学习姿势的现有数据集在姿势多样性、对象遮挡和视点方面不够具有挑战性。这使得姿势标注过程相对简单，并限制了已在其上训练的模型的应用。为了处理更多的人体姿势变化，我们提出了细粒度层次姿势分类的概念，其中我们将姿势估计制定为分类任务，并提出了一个数据集 Yoga-82，用于具有 82 的大规模瑜伽姿势识别类。 Yoga-82 由复杂的姿势组成，可能无法进行精细的注释。为了解决这个问题，我们根据姿势的身体配置为瑜伽姿势提供分层标签。该数据集包含一个三层层次结构，包括身体位置、身体位置的变化和实际的姿势名称。我们展示了 Yoga-82 上最先进的卷积神经网络架构的分类精度。我们还提出了 DenseNet 的几个分层变体，以利用分层标签。

Human pose estimation is a well-known problem in computer vision that focuses on locating the positions of human joints. Existing datasets used for pose learning are observed to be insufficiently challenging in terms of pose diversity, object occlusion, and viewpoint variation. This makes the pose annotation process relatively straightforward, and limits the applicability of models trained on such datasets. To address a wider range of human pose variations, we propose the concept of fine-grained hierarchical pose classification, where we formulate pose estimation as a classification task, and introduce a dataset named Yoga-82 for large-scale yoga pose recognition with 82 categories. Yoga-82 consists of complex poses that may be difficult to annotate with fine-grained details. To address this issue, we provide hierarchical labels for yoga poses based on their bodily configurations. This dataset features a three-layer hierarchical structure, which includes body position, variations of body positions, and the actual pose names. We evaluate the classification accuracy of state-of-the-art convolutional neural network (CNN) architectures on Yoga-82. We also propose several hierarchical variants of DenseNet to leverage the hierarchical labels.

搜集汇总

背景与挑战

背景概述

Yoga-82是一个用于细粒度人体姿态估计的瑜伽姿势识别数据集，包含82个类别，旨在解决现有数据集在姿势多样性、遮挡和视点方面的局限性。该数据集采用三层层次标签结构（身体位置、变化和姿势名称），以支持复杂姿势的分类任务，并已用于评估卷积神经网络架构的性能。