SurfaceRecipe: Information-Efficient Mesh Surface Representation for Lightweight 3-D Shape Classification - Supplementary Code and Visualization Materials

Abstract

Supplementary code and materials for 'SurfaceRecipe: Information-Efficient Mesh Surface Representation for Lightweight 3-D Shape Classification' (submitted to Computers and Graphics). SurfaceRecipe is a lightweight, mesh-informed 3-D shape classifier organised around a continuous geometric evidence signal. It builds fixed-length surface primitives from area-weighted triangle samples (coordinates, normals, areas, radii, edge scales, and a compact patch code) and derives a continuous per-primitive saliency score that drives an evidence-aware sampler (SCSS) and conditions bounded residual modulation of neighbourhood messages (SCRM). The package contains the full training/evaluation scripts, the ablation and reproducibility harness for the 589,814-parameter matched backbone, and the SCSS saliency visualization outputs (SVG/PNG/PLY/CSV) referenced in the paper. The model uses 0.600M parameters and reaches 93.3% overall accuracy on ModelNet40 without pretraining, voting, or test-time augmentation. This is the anonymized version prepared for peer review. GPU computing resources were provided by the Huawei AI 100-Schools Program (Computing Power Support).