The creators of the GWHD: Global Wheat Head Dataset are working towards automating the wheat head counting process through Deep Learning. However, this task is visually intricate, given factors like plant overlap, wind-induced blurriness, and variations in appearance based on factors such as maturity, color, genotype, and orientation. The dataset itself is extensive, containing over 6000 images, each with dimensions of 1024x1024 pixels, encompassing more than 300,000 distinct wheat heads, each paired with its corresponding bounding boxes. These images were taken across 11 different countries and span 44 unique measurement sessions. A “measurement session” here means a set of images captured at a consistent location over a few hours.

Diverse ground-based phenotyping platforms and cameras captured these images at heights ranging from 1.8 m to 3 m above the ground. They utilized various camera setups, resulting in Ground Sampling Distances (GSDs) that range between 0.10 and 0.62 mm, along with different field of view angles. Although these images are acquired in a nadir-viewing direction, occasional geometric distortions may appear in some sub-datasets due to distinct camera lens characteristics.

The authors of the GWHD dataset emphasized the need to harmonize diverse sub-datasets. To achieve this, they conducted a manual inspection, removing images captured too early in the growth stage when wheat heads were not clearly visible and avoiding those acquired during head senescence to prevent overlapping. Object scale, which is crucial for object detection methods, was influenced by both the object’s size (in mm) and image resolution. While wheat head dimensions generally measure around 1.5 cm in diameter and 10 cm in length, the image resolution varied significantly across sub-datasets. To address this, images were rescaled using bilinear interpolation to maintain a more uniform resolution at the head level. To align with the requirements of deep learning algorithms, the authors of the dataset divided the original images into square patches. These patches were carefully sized at 1024 x 1024 pixels to minimize the possibility of wheat heads intersecting patch boundaries and being partially cropped. Each patch typically contained approximately 20 to 60 heads, with only a few heads extending beyond the patch boundaries. The number of patches generated from each original image ranged from 1 to 6, depending on the specific sub-dataset.

The authors applied interactive labeling to wheat heads, creating bounding boxes encompassing the entire head pixel area. Labeling posed challenges when heads were not clearly visible due to obstruction by leaves or other heads. Partially hidden heads were only labeled if at least one spikelet was visible, typically occurring in images from the early growth stage when heads did not fully emerge. Head overlap was more common in images from cameras with a wider field of view, particularly along the image borders with oblique view angles. In cases where the bounding box was too large to include the awns, it was limited to covering the head itself. Additionally, heads that were cropped at the image edges were labeled only if at least 30% of their basal part was visible. The initial labeling results underwent an independent review by two individuals unaffiliated with the sub-dataset institution. In cases where significant discrepancies between reviewers were identified, an additional round of labeling and reviewing was initiated. Approximately 20 individuals participated in this labeling process, promoting collaboration and repeated reviews to maintain a high level of accuracy and consistency across the sub-datasets.