Overview

Forest systems play an important role in mitigating anthropogenic climate change and regulating global climate. However, due to difficulties in collecting wild data and lack of forestry expertise, the availability of large-scale forest datasets is very limited. In this work, we establish a new virtual forest dataset named PUTree. Our goal is to create a larger, more photo-realistic and diverse dataset as a powerful training resource in the wild forest. Early experimental results demonstrate its validity as a new forest benchmark for the evaluation of tree detection and segmentation algorithms, and its potential in broad application scenarios.

Our dataset can be utilized for both object-level and scene-level representation learning. The algorithms can take either single or multi-modality (RGB and depth) and single or multiple viewpoints (image or video frames) as input. The dataset and simulation tool can also be used to enable other forest applications such as automatic DBH measurement, 3D forest mapping and reconstruction, etc.