Horse10

姿态估计是测量行为的重要工具，因此广泛应用于技术、医学和生物学。由于深度学习算法和大规模数据集的创新，对人类的姿势估计已经变得非常强大。然而，典型的人体姿势估计基准，例如 MPII 姿势和 COCO，在不同的上下文中包含许多不同的个体 (>10K)，但每个个体只有很少的示例姿势。在姿态估计的实际应用中，用户希望通过仅在一小部分个体上标记几百帧来估计用户定义的身体部位的位置，但希望将其推广到新个体。因此，人们自然会提出以下问题：假设您已经训练了一种算法，该算法可以在给定（个体）动物的整个运动曲目中以高精度执行 - 它对外观略有不同或显着不同的不同个体的泛化效果如何?与这里常见的人体姿势估计基准不同，该设置是数据集每个人有许多（带注释的）姿势（> 200），但只有少数人（1-25）。
为了让该领域能够应对这一挑战，我们开发了一个名为 Horse-10 的新基准，由 30 匹不同的纯种马组成，其中 22 个身体部位由专家在 8,114 帧中标记。马有不同的毛色，在各种纯种一岁马销售和农场收集的数据的“野外”方面增加了额外的复杂性。

Pose estimation is a critical tool for behavioral measurement, and thus is widely applied in technology, medicine, and biology. Thanks to innovations in deep learning algorithms and large-scale datasets, human pose estimation has become remarkably robust. However, typical human pose estimation benchmarks, such as MPII Pose and COCO, contain over 10,000 distinct individuals across diverse contexts, but only a very limited number of pose examples per individual. In real-world applications of pose estimation, users aim to estimate the positions of user-defined body parts by labeling only a few hundred frames across a small subset of individuals, while intending to generalize to new unseen individuals. Thus, a natural question arises: suppose one has trained an algorithm that performs with high accuracy across the entire motion repertoire of a given (individual) animal—how well will it generalize to distinct individuals that look slightly or significantly different? Unlike common human pose estimation benchmarks, this setting features datasets where each individual has a large number of annotated poses (>200), but only a small number of individuals (1-25) are available.

To enable the field to address this challenge, we developed a novel benchmark named Horse-10, which consists of 30 distinct thoroughbred horses, with 22 body parts annotated by experts across 8,114 frames. Horses exhibit diverse coat colors, and the data was collected from various in-the-wild scenarios including thoroughbred yearling sales and farms, adding additional complexity to the benchmark.

搜集汇总

背景与挑战

背景概述

Horse10是一个专为动物姿态估计设计的基准数据集，包含30匹不同纯种马的22个身体部位标注，总计8,114帧图像，用于研究从少数个体到新个体的泛化能力。该数据集由多所大学和研究机构于2020年发布，数据规模为863.1MB，采用CC BY-NC 4.0许可证。