Segmentation of 3d plant point clouds

This thesis introduces a novel methodology for the semantic segmentation of 3D plant point clouds by projecting them into the 2D domain for efficient processing. The proposed multi-stage pipeline integrates grid downsampling, non-linear dimensionality reduction using t-SNE, binary image conversion, segmentation with 2D semantic segmentation architectures, and back-projection of predicted labels into 3D space. Extensive experiments were performed on four publicly available and diverse datasets: PLANesT-3D, Pheno4D, Soybean-MVS, and Plant3D. Multiple segmentation architectures were evaluated, including SegNet, U-Net, SegFormer, and Mask2Former. Results consistently showed that transformer-based architectures, particularly Mask2Former, provided superior accuracy and robustness across species and datasets. Importantly, this work established new state-of-the-art performance benchmarks on specific species. On the PLANesT-3D dataset, Mask2Former achieved an mIoU of 95.6% on pepper and 92.3% on rose, surpassing all previously reported methods. Likewise, on the Plant3D dataset, the proposed method achieved 95.47% mIoU on tobacco and 93.90% mIoU on tomato, setting new state-of-the-art results. The results underline both the methodological and practical significance of the proposed approach. By simplifying the segmentation task to the 2D domain and leveraging cutting-edge architectures, this research provides a scalable, accurate, and efficient framework. From an agricultural perspective, the improved segmentation of plant organs such as leaves and stems enhances the extraction of critical phenotypic traits, including leaf area, stem diameter, and overall plant architecture.