The project investigates if the use of relatively simple methods to generate synthetic datasets in the are of weed detection can yield comparable results to training models on conventionally annotated datasets. The problem analysis considers different datasets to base the work on, and explores the structure of the chosen dataset to then later use this structure information when generation several different synthetic datasets using a Cut, Paste and Learn approach by Dwibedi et al. The problem analysis also briefly discusses the choice of segmentation model for testing.
Following the analysis, the Design and Implementation of the dataset generation and segmentation model is described. Testing and results then describe how the methods explained are applied to generate different datasets with different blending techniques, and present the results, where surprisingly fully synthetic datasets outperformed the conventionally annotated dataset.