Using Simulated Data for Deep-Learning Based Real-World Apple Detection

Object detection using CNNs requires a large amount of data to achieve decent performance in real-world scenarios. The creation of traditional datasets involves acquiring numerous images and manually annotating them. In this paper, we introduce a method for simulating apple orchards utilizing the Unity 3D engine. We created a tool that uses this simulator to generate fully bounding-box annotated (simulated) datasets. We trained YOLOv5 models of different sizes on simulated data, real-world data, and a combination of both, and later tested the models on a real-world dataset to evaluate the suitability of our generated dataset. Our experiments show that object detection models trained on simulated data can achieve results on real-world images that are very similar to results of models trained solely on real-world data. We demonstrate that simulated data has the potential to eliminate the need for real-world datasets, thus saving a substantial amount of time. In this research, we focused our tests on a real-world dataset acquired under controlled settings, future work can be dedicated to evaluate the generalization ability of models trained on simulated datasets on more challenging real-world datasets.