Data Acquisition for Forestry Scenes

Abstract

Large quantities of samples are needed to train and benchmark today's data-driven machine-learning algorithms. Such well-established datasets for forestry scenes are scarce. This bachelor's thesis presents a simple, well-reported, and annotated forestry scene dataset for signal processing applications such as image segmentation and depth completion. We collected 5181 stereo images and 3D point clouds from three different locations and annotated a total of 200 images. A total of 3011 objects were manually annotated of which 1686 are tree trunks categorized into three instances: "Spruce", "Birch", and "Pine". In addition, a description of the data collection process is presented. This dataset will be later available on GitHub at: \\ \url{https://github.com/juanb09111/FinnForest}