PhoCaL

对象姿势估计对于机器人应用和增强现实至关重要。为了向社区提供具有高质量地面真相注释的基准，我们引入了一个用于类别级对象姿势估计的多模态数据集，该对象具有光度挑战性，称为phocalal。PhoCaL包括60个高质量的家庭对象的3D模型，超过8个类别，包括高反射，透明和对称对象。我们开发了一种新颖的机器人支持的多模态 (RGB，深度，极化) 数据采集和注释过程。它确保了不透明纹理，闪亮和透明物体的姿势的亚毫米精度，没有运动模糊和完美的相机同步。

Object pose estimation is critical for robotic applications and augmented reality. To provide the community with a benchmark equipped with high-quality ground-truth annotations, we introduce a multimodal dataset for category-level object pose estimation, which consists of photometrically challenging objects and is named PhoCaL. PhoCaL includes 3D models of 60 high-quality household objects spanning over 8 categories, encompassing highly reflective, transparent, and symmetric objects. We have developed a novel robot-supported multimodal (RGB, depth, polarization) data acquisition and annotation pipeline. This pipeline ensures sub-millimeter accuracy for the poses of opaque-textured, shiny, and transparent objects, free of motion blur and with perfect camera synchronization.

搜集汇总

背景与挑战

背景概述

PhoCaL是一个用于类别级对象姿势估计的多模态数据集，包含60个高质量家庭对象的3D模型，涵盖8个类别，特别关注高反射、透明和对称对象。该数据集采用创新的机器人支持多模态数据采集方法，确保亚毫米精度的姿势标注，适用于机器人应用和增强现实研究。