Cornell

该数据集由康奈尔大学计算机科学系的Yun Jiang、Stephen Moseson和Ashutosh Saxena创建，旨在通过RGBD图像（彩色图像与深度图）估计机械臂抓取物体时的7维夹爪配置（3D位置、3D方向和夹爪开口宽度）。数据集包含194张用于训练的物体图像（如马克笔、马提尼杯、橡胶玩具等）和128张用于测试的图像，涵盖9个物体类别。每个图像中手动标注了1至3个正确的抓取矩形，并随机生成了5个错误矩形作为负样本。数据集的创建过程基于监督学习，通过SVM排序算法训练模型，以预测抓取矩形的排名。该数据集的应用领域主要集中在机器人抓取任务，尤其是针对未见过的新物体的抓取。通过学习抓取矩形的表示，该数据集能够帮助机器人在复杂环境中实现更高效、更准确的抓取动作。

This dataset was created by Yun Jiang, Stephen Moseson, and Ashutosh Saxena from the Department of Computer Science at Cornell University, with the goal of estimating the 7-dimensional gripper configuration (including 3D position, 3D orientation, and gripper opening width) of a robotic arm during object grasping using RGBD images (color images paired with depth maps). The dataset consists of 194 training object images (e.g., markers, martini glasses, rubber toys, etc.) and 128 test images, spanning 9 object categories. Each image is manually annotated with 1 to 3 positive grasp rectangles, while 5 randomly generated incorrect rectangles are included as negative samples. This dataset is designed for supervised learning tasks, where an SVM ranking algorithm is utilized to train models for predicting the ranking of grasp rectangles. Its primary application domains are robotic grasping tasks, particularly grasping previously unseen novel objects. By learning the representations of grasp rectangles, this dataset enables robots to perform more efficient and accurate grasping operations in complex environments.