PixelCropRobot dataset: images of vegetables crops in different phenological stages taken in greenhouses

Dataset created under the PixelCropRobot project, developed by FCUP, INESC TEC and FEUP.

Dataset folder:

This folder contains the images of each species in two formats (3456 × 4608 pixels and 864 × 1152 pixels), the annotations of the 864 × 1152 px. images, in Pascal VOC (.xml) and YOLO (.txt) formats and also a set of Python scripts useful for managing the dataset.

The aim was to capture images of eight crops selected taking into account: i) the short cultural cycle, ii) low intensity of cultural operations and low pressure of pests and diseases and also iii) wide use in Portuguese cuisine.

The images were captured using a smartphone (Huawei Mate 10 Lite), with 16 megapixels (MP) resolution (3456 × 4608 px.), in Professional mode (no flash, continuous autofocus, automatic ISO and shutter speed). Image collection took place at different hours of the day, with variable lighting conditions.

The images are divided as follows (in parenthesis are the classes):

Arugula - 312 (coty, minus9, plus9)

Carrot - 533 (coty, smallleaves, carrot)

Coriander - 323 (coty, smallleaves, coriander)

Lettuce - 1426 (coty, minus9, plus9, ready)

Radish - 494 (coty, smallleaves, bigleaves, root)

Spinach - 270 (spinach, big)

Swiss chard - 454 (coty, chard)

Turnip - 313 (coty, smallleaves, turnip)

To standardise the dataset, each image was renamed according to the corresponding EPPO (European and Mediterranean Plant Protection Organization) code and the date of creation of that image. The size of each image was also reduced four times (to 864 × 1152 pixels) to facilitate processing. For example, an image of lettuce captured on June 22 presents the name as follows: LACSA_Jun_22_x_864_1152.jpg.

3D:

This folder contains 5 sets of 2D images - coriander, grown, plots, radish and smalls, the corresponding sparse and dense points clouds and the 3D model. Also has the corresponding Agisoft Metashape worksheet.

There are several processes, based on photogrammetry techniques, applied for the three-dimensional reconstruction of plants: i) Structure from Motion (SfM) - generates a three-dimensional point cloud representing the plant in a dispersed way, ii) Multi-View Stereo (MVS), applying the parameters calculated by the previous process, a dense point cloud is generated and iii) surface reconstruction, the transformation of the dense point cloud into a mesh of surfaces.

The software Agisoft Metashape was used to calculate the SfM and MVS values and therefore to create the point clouds. To reconstruct the plots, not only was the Agisoft Metashape software applied, but also OpenDroneMap software, a set of open source tools for processing images into point clouds, georeferenced textured mesh and orthophoto GeoTIFF.