HiMamba-Net: a Hilbert-serialized Mamba network for 3D point cloud instance segmentation

Area of Science:

• Computer Vision
• Robotics
• Agricultural Technology

Background:

• Instance segmentation of 3D point clouds is challenging due to complex structures and occlusions.
• Existing methods struggle with large-scale agricultural scenes featuring dense, irregular, and repetitive plant geometries.

Purpose of the Study:

• To propose HiMamba-Net, an efficient framework for spatially coherent instance segmentation of agricultural point clouds.
• To address limitations of current methods in handling the unique challenges of field-scale agricultural data.

Main Methods:

• Utilizes Hilbert space-filling curves for spatial serialization, preserving geometric locality.
• Employs selective state space models (SAMamba blocks) for linear-complexity sequential modeling.
• Integrates patch-based feature extraction, multi-scale graph context aggregation, and multi-task learning.

Main Results:

• HiMamba-Net achieved 91.8% mIoU for semantic segmentation and 91.9% mAP for instance segmentation on the Crops3D dataset.
• Outperformed baseline methods in accuracy and efficiency for agricultural point cloud segmentation.
• Ablation studies confirmed the effectiveness of Hilbert serialization and selective state space modeling.

Conclusions:

• Spatially coherent serialization combined with linear-complexity sequence modeling offers an effective solution for large-scale 3D point cloud instance segmentation.
• HiMamba-Net demonstrates strong performance in complex agricultural environments.
• The proposed framework advances the state-of-the-art in agricultural point cloud analysis.