Synthetic Data for Agriculture

Project description

How can domain-adapted generative models be effectively used to create synthetic field images for training robust multi-class agricultural object detectors, and how does this approach compare to traditional data collection methods?

Context

Context:
This research supports precision farming for Luxeed Robotics, developing vision models to distinguish crops (onions) from weeds for laser-weeding systems.

Challenge:
The primary goal is to generate diverse, large-scale datasets synthetically to avoid the high cost and labor of real-world data collection.

Results

This study demonstrates that while purely synthetic data is insufficient for training robust agricultural object detectors, a hybrid approach offers a powerful solution. By augmenting a set of 101 real images with 200 targeted synthetic images, the resulting RF-DETR model achieved a 67.1% mAP, surpassing models trained on either data source alone or on a larger hybrid set

About the project group

The project focuses on research on synthetic data for computer vision and specifically for agriculture, where generated images are used to train an object detection model.