Tree trunk detection in urban scenes using a multiscale attention-based deep learning method
Rao Li, Guodong Sun, Sheng Wang, Tianzhuzi Tan, Fu Xu
Abstract
Precise identification of tree trunks contributes to the understanding of urban green dynamics. Previous attempts to develop tree trunk detection methods have faced limitations in respect of precision and generalization due to the use of hand-engineered features and the constraint of single-species detection. In this study, we construct a new tree trunk dataset considering the object’s strong diversity and propose a deep model to detect and segment the salient tree trunks or even branches in urban scenes. Comprehensive experiments are performed to evaluate our model. The presented method exhibits exceptional performance, evidenced by its outstanding scores across seven evaluation metrics, indicating its capability to segment tree trunks of different species, even if they exhibit significant variations in appearance. Specifically, our model demonstrates outstanding accuracy in detecting trunks with intricate furcations, as well as effectively identifying trunks that are partially occluded.