SPREAD: A large-scale, high-fidelity synthetic dataset for multiple forest vision tasks

Abstract

We present the Synthetic Photo-realistic Arboreal Dataset (SPREAD), a state-of-the-art synthetic dataset specifically designed for forest-related machine learning tasks. Developed using Unreal Engine 5, SPREAD goes beyond existing synthetic forest datasets in terms of realism, diversity, and comprehensiveness. It includes RGB, depth images, point clouds, semantic and instance segmentation labels, along with key parameters such as tree ID, location, diameter at breast height (DBH), height, and canopy diameter. In exemplary experiments, we found that SPREAD significantly reduces the need to use real-world datasets for trunk segmentation tasks and enhances model segmentation performance. Specifically, by pretraining on SPREAD, MobileNetV3 and DeepLabV3 models require only 25% of a fine-tuning real-world dataset to match or even surpass the performance of ImageNet-pretrained models fine-tuned on the entire real-world dataset. Furthermore, our hybrid training experiments demonstrate that by combining SPREAD and real data at a 1:1 or 2:1 ratio greatly improves task performance. For the canopy instance segmentation task, SPREAD pretraining still provides varying degrees of performance improvement for the models. All datasets, data collection frameworks, and codes are available at https://github.com/FrankFeng-23/SPREAD.