Data for publication. CANOPIES Grape bunch and peduncle dataset

In September 2021, data of grape bunches and peduncles were collected in the vineyard of the Corsira Agricultural Cooperative Society (Aprilia, Italy) (Via Corsira (https://goo.gl/maps/UdqVTNQnJTGzRnY39), where table grapes are grown. A total of 45 sequences were recorded, with a total duration of 3,660 seconds (61 minutes) and containing a total of 103,775 images and 80,489 point clouds. The images were acquired on different days and at different times, from 9am to 5pm. They had different lighting conditions due to weather conditions, density of foliage, or time of day. The images were also taken at different distances from the bunches, ranging from 30cm to 2m. 

From the data recorded, ten sequences were selected in which the peduncles and bunches were fully or partially visible, discarding those taken from too far away to identify the peduncle. The CANOPIES Grape Bunch and Peduncle (GBPD) dataset contains a total of 810 RGB images and with their correspondent binary masks for grape bunches and peduncles.

The images were recorded using an Intel RealSense Depth Camera D435. Both 640x480 and 1280x720 resolutions were used.

The dataset was labelled using VGG Image Annotator (VIA) [1], a free software that can be used without any installation. The labelling work was divided among 8 people. Each of them was given strict guidelines on how to label the images in order to keep the labelling consistent. 

These data were used for accurate detection of table grapes and peduncles for harvesting which methods can be seen in the article [2].

 References

[1] A. Dutta, A. Gupta and A. Zisserman (2019), VGG image annotator (VIA). http://www.robots.ox.ac.uk/~vgg/software/via/

[2] Gabriel Coll-Ribes, Ivan J. Torres-Rodriguez, Antoni Grau, Edmundo Guerra, Alberto Sanfeliu (2023), Accurate detection and depth estimation of table grapes and peduncles for robot harvesting, combining monocular depth estimation and CNN methods. Computer and Electronics in Agriculture. Vol. 215, December 2023, 108362. https://doi.org/10.1016/j.compag.2023.108362

2021年9月，研究团队于意大利阿普里利亚（Aprilia）的科尔西拉农业合作社（Corsira Agricultural Cooperative Society）葡萄园（地址：Via Corsira，地图链接：https://goo.gl/maps/UdqVTNQnJTGzRnY39）采集了葡萄果串与果柄数据，该园区种植鲜食葡萄。本次共记录45段序列，总时长3660秒（合61分钟），包含总计103775张图像与80489个点云。

图像采集于不同日期与时段（上午9时至下午5时），受天气状况、枝叶密度以及拍摄时段影响，光照条件存在差异；拍摄距离也各不相同，范围为30厘米至2米。

从所记录的数据中，研究团队筛选出10段序列，其中果柄与果串完全或部分可见，剔除了因拍摄过远而无法识别果柄的序列。CANOPIES葡萄果串与果柄（GBPD）数据集包含总计810张RGB图像，以及对应葡萄果串与果柄的二值掩码。

图像采集使用英特尔实感深度相机D435（Intel RealSense Depth Camera D435），支持640×480与1280×720两种分辨率。

本数据集使用VGG图像标注器（VGG Image Annotator, VIA）[1]进行标注，该软件为免费开源工具，无需安装即可使用。标注工作由8名人员分工完成，为保证标注一致性，每名参与者均收到了严格的图像标注规范指南。

该数据集被用于鲜食葡萄与果柄的精准检测，以服务于机器人采收任务，相关方法详见文献[2]。

参考文献
[1] A. Dutta, A. Gupta 与 A. Zisserman (2019), VGG图像标注器(VIA)。http://www.robots.ox.ac.uk/~vgg/software/via/
[2] Gabriel Coll-Ribes、Ivan J. Torres-Rodriguez、Antoni Grau、Edmundo Guerra、Alberto Sanfeliu (2023), 结合单目深度估计与卷积神经网络方法实现机器人采收用鲜食葡萄与果柄的精准检测及深度估计。《农业与计算机电子学》（Computer and Electronics in Agriculture），第215卷，2023年12月，文章编号108362。https://doi.org/10.1016/j.compag.2023.108362