The ACRE Crop-Weed Dataset for Benchmarking Weed Detection Models on Maize and Beans Fields

The increasing pressure on agricultural productivity due to population growth and climate change could be relieved by the integration of robotics into agriculture. However, developing autonomous robots for real-world agricultural scenarios is challenging due to environmental variability and the need for representative datasets. This study introduces the ACRE Crop-Weed Dataset, designed for benchmarking weed detection models on maize and bean fields. The dataset, collected during the 1st ACRE Field Campaign, consists of 1000 RGB images captured in June 2022, featuring maize, bean crops, and four weed types. Manual instance segmentation annotations are included. Various experiments were conducted using the ACRE dataset, such as training a single YOLOv8 model on the entire dataset and separate models for maize and bean crops to assess performance differences. Generalization capabilities were evaluated by training on datasets from previous years and testing on the ACRE dataset. The findings contribute to developing efficient weeding robots for precision agriculture, addressing environmental variability and the need for representative datasets. The dataset can be downloaded at https://doi.org/10.5281/zenodo.8102217