PoseBench
收藏资源简介:
PoseBench是由悉尼大学、京东探索研究院和南洋理工大学共同开发的综合性基准数据集,旨在评估姿态估计模型在面对真实世界损坏时的鲁棒性。该数据集涵盖了人类和动物姿态估计,包括60种代表性模型,涉及多种损坏类型如模糊、噪声、压缩和颜色损失等。创建过程中,数据集通过应用四类损坏到原始验证集上,确保了评估的全面性和公平性。PoseBench的应用领域广泛,包括人机交互、具身AI和自动驾驶等,旨在解决现有模型在实际部署中因数据损坏而引发的安全风险问题。
PoseBench is a comprehensive benchmark dataset jointly developed by the University of Sydney, JD Explore Academy, and Nanyang Technological University, which aims to evaluate the robustness of pose estimation models against real-world corruptions. This dataset covers human and animal pose estimation tasks, includes 60 representative models, and involves various corruption types such as blur, noise, compression, and color loss. During the dataset creation process, four types of corruptions were applied to the original validation set to ensure the comprehensiveness and fairness of the evaluation. PoseBench has a wide range of application fields including human-computer interaction, embodied AI, autonomous driving and others, and it is designed to solve the security risks caused by data corruption when existing models are deployed in real-world scenarios.




