ProPose
收藏资源简介:
ProPose是由昆士兰大学、阿德莱德大学及澳门城市大学联合构建的大规模人体姿态估计基准,旨在统一完整肢体、残肢与假肢的拓扑表示。该数据集包含36,400张图像、63,000个实例,覆盖生物、假肢及缺失关节的语义标注,并引入6个远端关键点(指尖、脚跟、脚趾)以完整刻画运动链。针对假肢样本的稀缺性与长尾分布,研究者设计了真实到合成的数据扩充管线,通过大语言模型合成9,558张新图像,使假肢关节比例提升数十倍。ProPose致力于解决现有基准对肢体差异人群的表示偏见,为包容性姿态估计及人机交互研究奠定基础。
ProPose is a large-scale human pose estimation benchmark jointly developed by The University of Queensland, The University of Adelaide, and City University of Macau, which aims to unify the topological representations of intact limbs, residual limbs and prosthetics. This dataset contains 36,400 images and 63,000 instances, with semantic annotations for biological, prosthetic and missing joints, and introduces 6 distal keypoints (fingertips, heels, toes) to fully characterize the kinematic chain. To address the scarcity and long-tail distribution of prosthetic samples, the researchers designed a real-to-synthetic data augmentation pipeline, which synthesized 9,558 new images via large language models (LLMs), boosting the proportion of prosthetic joints by dozens of times. ProPose is committed to addressing the representation bias of existing benchmarks towards people with limb differences, laying a foundation for inclusive pose estimation and human-computer interaction research.

- 1Topology-Unified 2D Pose Estimation across Intact, Residual and Prosthetic Limbs昆士兰大学; 阿德莱德大学·澳大利亚机器学习研究所; 澳门城市大学 · 2026年




