Reproduce “PLANesT-3D: A new annotated dataset for segmentation of 3D plant point clouds” on OpenResearch

About this paper

Kerem Mertoğlu, Yusuf Şalk, Server Karahan Sarıkaya, Kaya Turgut, Yasemin Evrenesoğlu, Hakan Çevikalp, Ömer Nezih Gerek, Helin Dutağacı and 1 more

Researchers introduced PLANesT-3D, a new public dataset of 34 annotated 3D color point clouds for pepper, rose, and ribes plants, acquired using Structure from Motion/Multi-view Stereo. The study also evaluated a novel semantic segmentation method, SP-LSCnet, which performed comparably to or slightly better than PointNet++, while RoseSegNet achieved the highest mean IoU of up to 95.5% on the dataset.

Key insights

• SfM/MVS provides a low-cost method for 3D plant data acquisition, but introduces challenges like varying point density, reconstruction noise, and occlusions, which are crucial for training algorithms that generalize to real-world conditions.
• The two-stage, modular design of SP-LSCnet, with its t-SNE-based superpoint extraction, offers increased interpretability into the segmentation process compared to 'black-box' end-to-end deep learning models.
• Species diversity and a larger number of distinct plant instances in a 3D dataset are essential for evaluating the robustness and transferability of machine learning models across different plant architectures and reducing the need for extensive manual annotation.

Summary by alphaXiv.