A Self-Structure-Enhanced Algorithm for Pig Point Cloud Completion

Area of Science:

• Computer Vision
• Animal Science
• 3D Reconstruction

Background:

• Three-dimensional phenotypic data are vital for pig growth, health, and intelligent livestock management.
• Acquiring complete 3D point clouds in real environments is challenging due to occlusion, sensor noise, and limited viewpoints.

Purpose of the Study:

• To develop a self-structure-enhanced completion method for addressing local missing structures and geometric discontinuities in pig point clouds.
• To introduce a visibility-incompleteness mask (VIM) to differentiate occlusion from true missingness and a prior-adaptive hybrid generation (PAHG) strategy.

Main Methods:

• A global-to-local two-stage framework: self-view fusion network (SVFNet) for coarse completion and self-structure dual generator (SDG) for refinement.
• Integration of incomplete point clouds with orthogonal self-projected depth maps.
• Explicit modeling of geometric missingness and multi-view visibility using VIM.

Main Results:

• The proposed method achieved competitive performance on a dataset of 1042 pig point cloud pairs.
• Key metrics include F-Score@1% of 0.653, CD-L1 of 9.766, and CD-L2 of 0.353.
• Demonstrated superior aggregate performance compared to state-of-the-art completion methods.

Conclusions:

• The self-structure-enhanced completion method effectively reconstructs incomplete 3D pig point clouds.
• The VIM and PAHG strategies enhance the accuracy and robustness of the completion process.
• This advancement supports precision feeding, genetic selection, and intelligent livestock management.