Generalization Approach for CNN-based Object Detection in Unconstrained Outdoor Environments

Recent advances in object detection using convolutional neural networks provide new opportunities to address various problems faced in agriculture. However, training a reliable object detection model which can be generalized to different environments requires large amounts of data which is difficult to obtain. This is particularly true when considering detection of wilding conifers, which are scattered across a diverse, unconstrained environment. Here we argue that a reliable dataset can be developed for wilding detection by synthetically enlarging the training dataset. A reliable detection model was trained using only 100 real images of wilding conifers.