real RGB-D videosynthetic PBR RGB-Dscanned object CAD modelspart-level articulated annotations
关键依赖
calibrated three-camera capture system for annotationobject scans and canonical alignmentsynthetic training dataGT detections for reported pose baselinesGT first pose for trackingGT masks / sizes for several tracking baselines
完整论文元数据作者、单位、研究方向、数据集、指标、方法标签和阅读状态展开收起
作者
Yang You, Kai Xiong, Zhening Yang, Zhengxiang Huang, Junwei Zhou, Ruoxi Shi, Zhou Fang, Adam W. Harley, Leonidas Guibas, Cewu Lu
作者单位
Stanford University / Shanghai Jiao Tong University / Horizon Robotics / UC San Diego
calibrated three-camera capture system for annotationobject scans and canonical alignmentsynthetic training dataGT detections for reported pose baselinesGT first pose for trackingGT masks / sizes for several tracking baselines
Reading Notes
粗读 / 精读
已生成 src/paper-notes/pace-eccv2024/coarse.md
论文粗读笔记
用于快速判断论文定位、问题设定、核心方法、实验结论和是否值得精读。
论文粗读笔记:PACE
1. 基本信息
标题:PACE: A Large-Scale Dataset with Pose Annotations in Cluttered Environments
作者:Yang You, Kai Xiong, Zhening Yang, Zhengxiang Huang, Junwei Zhou, Ruoxi Shi, Zhou Fang, Adam W. Harley, Leonidas Guibas, Cewu Lu
机构:Stanford University;Shanghai Jiao Tong University;Horizon Robotics;UC San Diego
作者在带官方 GT 的 NOCS-REAL275 Scene 1 上重新标注,得到平均 0.9° rotation / 2.3 mm translation 差异。这个验证支持局部准确性,但只覆盖一个较简单场景,不能完全量化 PACE 中 severe occlusion、快速运动和 articulated parts 的标注不确定性。
mask 由 CAD depth rendering 生成;出现手部交互时,先用 SAM 估 hand mask,再从 rendered object mask 中扣除。这会带来两个边界:pose GT 的小误差会变成边缘 mask error;SAM hand mask 的漏分 / 过分也会影响 visible ratio。
Real2Real All 仍比 Single 低 57.2–78.5 points;作者观察到训练 loss plateau / underfitting。因此不能把 PACE 的低分全部解释为渲染域差距。更大的模型、object-conditioned representation、retrieval / foundation features、balanced sampling 和 modular geometry solver 都可能是后续方向。