Learning to Reconstruct Botanical Trees from Single Images

Fig. 1. Single image tree reconstruction: our method automatically reconstructs trees from single images (a). We obtain semantic segmentation masks of trees to identify branches and leaves (b). We introduce Radial Bounding Volumes (c) as a fixed-size representation for 3D tree models and show that this representation can directly be learned with neural networks. We then use the predicted RBVs to automatically reconstruct tree models with a high degree of visual fidelity (d). To reconstruct multiple trees in a single image we detect bounding boxes (a, red) before obtaining the semantic segmentation masks.