3B bitki nokta bulutlarının 2B’de bölütlenmesi 2D segmentation of 3D plant point clouds

In this paper, we introduce a new approach for semantic segmentation of 3D plant point clouds through image-based analysis of their 2D mappings. The map points are obtained by t-SNE and are converted to 2D binary images. The natural clusters delivered by t-SNE are processed as separate images through deep neural networks that had been designed to perform semantic segmentation on 2D images. We tested our approach with SegNet and UNet, as two representatives of architectures for semantic segmentation. We generated multiple maps through multiple runs of t-SNE to increase the size of the training set, as well as to merge predictions during inference. The results on the publicly available PLANesT-3D data set in comparison with PointNet++ demonstrate that our approach is effective for semantic segmentation of 3D point clouds in the 2D image domain.