SemanPartmesh

Semantic-guided quadrilateral meshing pipeline integrating PartField (semantic features), NeurCross (cross field training), and quad_extract (MIQ + libQEx extraction).

Directory Layout

SemanPartmesh/
├── run_pipeline.py          # Main pipeline entry (PartField → NeurCross → extract)
├── extract_quad.py          # Stage 3: quad mesh extraction wrapper
├── build_complexity_map.py  # Semantic complexity / size field utilities
├── visualize_feature_clusters.py
├── requirements.txt
│
├── scripts/                 # Shell launchers and environment setup
│   ├── install_env.sh
│   ├── run_pipeline_simplified.sh
│   ├── run_baseline*.sh
│   └── ...
│
├── eval/                    # Evaluation metrics (angle distortion, Jacobian, etc.)
├── PartField/               # PartField source (vendored)
├── NeurCross/               # NeurCross source (vendored)
├── quad_extract/            # C++ quad extraction (CMake)
│
├── experiments/             # Historical experiment outputs
│   ├── compare/             # Single-mesh baseline vs pipeline comparisons
│   ├── reconstruction/      # Reconstruction dataset batch runs
│   ├── hunyuan/             # Hunyuan patch extraction script + results
│   │   └── run_hunyuan_patch_extract.py
│   ├── pipeline/            # Standard pipeline outputs
│   ├── baseline/            # Baseline-only outputs (created at runtime)
│   └── cheburashka/         # Complexity analysis outputs
│
└── docs/
    └── 脚本说明.md          # Script usage guide (Chinese)

Quick Start

# Install environment
bash scripts/install_env.sh

# Run simplified pipeline on a mesh or directory
bash scripts/run_pipeline_simplified.sh input/armadillo.obj

# Run baseline (NeurCross only, no PartField guidance)
bash scripts/run_baseline_simplified.sh input/

Pipeline Stages

1. PartField — extract semantic features (partfield conda env)
2. NeurCross — train cross field with optional semantic guidance (neurcross conda env)
3. extract_quad.py — MIQ parametrization + libQEx quad mesh generation

External Dependencies (not in repo)

These must be set up separately before running the full pipeline:

Path	Purpose
PartField/model/model_objaverse.ckpt	PartField checkpoint
libigl/, libQEx/	C++ dependencies for quad_extract
Baseline/NeurCross/	Baseline training scripts
instruction_guidance/	Instruction-guided mode dataset + metadata
input/	Default input mesh directory

Output Location

New pipeline runs write to experiments/pipeline/pipeline_output/ by default (with optional timestamp suffix). Override with --output_dir.

See also: 脚本说明（中文） — detailed guide for all scripts and workflows.