Global Wheat Head Detection 2021: An Improved Dataset for Benchmarking Wheat Head Detection Methods

The Global Wheat Head Detection (GWHD) dataset was created in 2020 and has assembled 193,634 labelled wheat heads from 4700 RGB images acquired from various acquisition platforms and 7 countries/institutions. With an associated competition hosted in Kaggle, GWHD_2020 has successfully attracted attention from both the computer vision and agricultural science communities. From this first experience, a few avenues for improvements have been identified regarding data size, head diversity, and label reliability. To address these issues, the 2020 dataset has been reexamined, relabeled, and complemented by adding 1722 images from 5 additional countries, allowing for 81,553 additional wheat heads. We now release in 2021 a new version of the Global Wheat Head Detection dataset, which is bigger, more diverse, and less noisy than the GWHD_2020 version.

Automated summary

The Global Wheat Head Detection dataset created in 2020 attracted attention from computer vision and agricultural science communities. The 2021 version of the dataset is larger, more diverse, and less noisy than the previous version after improvements and updates.