Reproduce “OmniPlantSeg: Species Agnostic 3D Point Cloud Organ Segmentation for High-Resolution Plant Phenotyping Across Modalities” on OpenResearch

About this paper

Andreas Gilson, Lukas Meyer, Oliver Scholz, Ute Schmid

The OmniPlantSeg pipeline presents a species- and modality-agnostic approach for 3D plant organ segmentation, utilizing a novel resolution-retaining sub-sampling algorithm (KD-SS) to process high-resolution point clouds. It achieved a mean IoU of 95.6% for Pepper plants and 96.3% accuracy for Sorghum, demonstrating robust performance across diverse plant species and sensor modalities.

Key insights

• Retaining the full resolution of point cloud data through an efficient sub-sampling method like KD-SS is crucial for capturing fine-grained details necessary for high-accuracy plant phenotyping.
• A unified pre-processing pipeline integrated with a suitable deep learning model can effectively generalize 3D plant organ segmentation across a wide range of plant species and diverse sensor modalities.
• Integrating normalized coordinates per sub-sphere as additional features enhances segmentation performance, particularly for data from sensors like LiDAR which may have coarser intrinsic resolution.

Summary by alphaXiv.