Shelf&Tote Benchmark Dataset (MIT-Princeton Amazon Picking Challenge 2016 Shelf&Tote Benchmark Dataset)

近年来，仓库自动化引起了极大的兴趣，其中最明显的可能是亚马逊拣货挑战赛 (APC)。实现完全自主的取放系统需要一个强大的视觉系统，能够可靠地识别物体及其 6D 姿势。但是，由于环境杂乱、自闭塞、传感器噪声和种类繁多的物体，解决方案避开了仓库设置。在本文中，我们展示了一个视觉系统，该系统在 APC 2016 的装载和拣选任务中分别获得第三和第四名。我们的方法利用多视图 RGB-D 数据和数据驱动的自我监督学习来克服上述困难。更具体地说，我们首先使用完全卷积神经网络对场景的多个视图进行分割和标记，然后将预扫描的 3D 对象模型拟合到生成的分割中以获得 6D 对象姿势。训练用于分割的深度神经网络通常需要大量带有手动标签的训练数据。我们提出了一种自我监督的方法来生成大型标记数据集，而无需繁琐的手动分割，可以轻松扩展到更多对象类别。我们证明了我们的系统可以在各种场景下可靠地估计物体的 6D 姿态。

In recent years, warehouse automation has garnered tremendous interest, most notably perhaps the Amazon Picking Challenge (APC). Achieving fully autonomous pick-and-place systems requires a robust vision system that can reliably identify objects and their 6D poses. However, solutions have eluded warehouse settings due to cluttered environments, occlusions, sensor noise, and the wide variety of objects. In this paper, we present a vision system that achieved 3rd and 4th place in the loading and picking tasks of APC 2016, respectively. Our method leverages multi-view RGB-D data and data-driven self-supervised learning to overcome the aforementioned challenges. More specifically, we first use fully convolutional neural networks to segment and label multiple views of the scene, then fit pre-scanned 3D object models to the resulting segmentations to obtain the 6D object poses. Training deep neural networks for segmentation typically requires large amounts of training data with manual annotations. We propose a self-supervised approach to generate large labeled datasets without tedious manual segmentation, which can be easily scaled to more object categories. We demonstrate that our system can reliably estimate the 6D poses of objects across various scenarios.

搜集汇总

背景与挑战

背景概述

Shelf&Tote Benchmark Dataset是一个用于仓库自动化视觉系统的数据集，特别针对亚马逊拣货挑战赛（APC）中的6D姿态估计问题。该数据集利用多视图RGB-D数据和自我监督学习方法，旨在解决环境杂乱、自闭塞和传感器噪声等挑战。