Abstract

This paper presents a 2D pose estimation solution to the bin-picking problem for robotic grasping systems. By extending a pretrained object detection model, namely DETR, with pose and visibility prediction heads we obtain classification, center, 2D rotation and occlusion scores for every detected object. The augmented model is trained and evaluated on synthetically generated images representing the real environment for faster and more flexible acquisition of data. The results show an average angle error of 3.23 deg for cylindrical and cuboid shape objects.

中文翻译：

---

使用平面姿态估计功能增强预训练的物体检测模型

摘要

本文提出了一种针对机器人抓取系统的箱子拾取问题的 2D 位姿估计解决方案。通过使用姿态和可见性预测头扩展预训练的对象检测模型（即 DETR），我们获得每个检测到的对象的分类、中心、2D 旋转和遮挡分数。增强模型在代表真实环境的合成生成图像上进行训练和评估，以便更快、更灵活地获取数据。结果显示，圆柱形和长方体形状物体的平均角度误差为 3.23 度。