GaussianPlant Dataset

We created a new real-world benchmark dataset for 3D plant modeling and structural reconstruction, named Gaussianplant dataset. The dataset contains 10 artificial plants all captured under indoor lighting conditions. We use a variety of plant species, including those with large leaves (e.g., Umbellata) and small leaves (e.g., Benjamina), where the characteristics of shape, structure, and occlusion differ.For each scene, we acquire more than 120 multi-view images and reconstruct the plants using multi-view stereo, to create a dense and reliable point cloud. During the capture, we carefully plan viewpoints to minimize occlusions, ensuring that the reconstructed point clouds are as complete as possible.The reconstructed point clouds are then manually cleaned to remove background and floating noise, which we use as the ground-truth shapes. On top of these point clouds, we annotate two types of ground-truth labels for benchmarking: (i) a binary leaf-branch segmentation and (ii) leaf instance labels.Specifically, we import each dense point cloud into Blender, and use its vertex selection tools to interactively assign labels to points. For plants with extremely dense foliage, such as Ash}and Benjamina, reliable per-leaf delineation is infeasible, so we only provide binary leaf-branch labels and exclude them from the leaf instance segmentation evaluation.