A-Occ-Plant: Plant occluded point cloud completion via amodal segmentation

Area of Science:

• Computer Vision
• Plant Science
• 3D Reconstruction

Background:

• 3D plant point clouds often have missing data due to self-occlusion, hindering analysis.
• Accurate 3D plant models are crucial for phenotypic trait extraction and reconstruction.

Purpose of the Study:

• Introduce A-Occ-Plant, a novel method for completing occluded plant point clouds.
• Improve the accuracy and efficiency of 3D plant analysis.

Main Methods:

• Convert 3D point clouds to images for 2D amodal segmentation and completion.
• Utilize a hierarchical Transformer with cross-scale attention for coarse-to-fine processing.
• Employ view-consistent Gaussian splatting to generate complete 3D point clouds from completed masks.

Main Results:

• A-Occ-Plant significantly improves upon state-of-the-art 3D point-cloud completion methods.
• Achieved 264.8% improvement in LPIPS and 8.3% gain in SSIM.
• Demonstrated effective recovery of phenotypic traits (leaf count, angle, height) and generalization to soybean.

Conclusions:

• A-Occ-Plant offers a robust solution for 3D plant point cloud completion, overcoming occlusion challenges.
• The method is computationally efficient, requiring fewer parameters and running faster than existing approaches.
• Enables more accurate plant phenotyping and 3D reconstruction for various crops.