📘 PASDF: Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation

Bozhong Zheng^1*, Jinye Gan^1*, Xiaohao Xu^2†, Xintao Chen¹, Wenqiao Li¹, Xiaonan Huang², Na Ni^1†, Yingna Wu^1†

¹ShanghaiTech University    ²University of Michigan, Ann Arbor

^*Equal contribution.   ^†corresponding author.

---

🌍 Overview

PASDF is a framework designed for 3D anomaly localization and geometric repair, leveraging high-quality continuous geometric representations via neural implicit functions.
It bridges 3D anomaly detection and surface reconstruction by learning a Signed Distance Function (SDF) representation of object geometry.

🧩 1. Installation

conda create -n PASDF python=3.10
conda activate PASDF
cd path/to/PASDF
bash install.sh PASDF

📦 2. Download Required Data

2.1 Datasets

Real3D-AD Dataset

• Training / Evaluation Data (PCD format)

  • Google Drive
  • Baidu Disk (code: vrmi)

• Raw Data (PLY format)

  • Google Drive
  • Baidu Disk (code: vvz1)

Anomaly-ShapeNet (ShapeNetAD) Dataset

• Training / Evaluation Data (PCD & OBJ format)

  • Baidu Disk (code: case)
  • Hugging Face Dataset Page

2.2 Preprocessed Templates and Model Weights

Before running inference (PASDF/Test/AD_test.py), download the preprocessed SDF samples and pretrained model weights.

🔗 Google Drive Folder

This folder contains:

data/
results/

📁 Directory Structure

Place the contents into your project as follows:

PASDF/
├── data/                           ← Template files for both datasets
│   ├── ShapeNetAD/                 ← Template meshes for ShapeNetAD
│   ├── Real3D_AD/                  ← Template meshes for Real3D-AD
│
├── results/                        ← Pretrained weights and preprocessed SDF data
│   ├── ShapeNetAD/
│   │   ├── runs_sdf/               ← Trained model weights (.pt files)
│   │   └── samples_dict_ShapeNetAD.npy   ← Preprocessed SDF samples
│   │
│   ├── Real3D_AD/
│   │   ├── runs_sdf/
│   │   └── samples_dict_Real3D_AD.npy
│
└── Test/
    ├── AD_test.py
    └── infer.py

⚙️ 3. Usage

🧱 3.1 Anomaly_ShapeNet(ShapeNetAD)

⚠️ Note: If you have already downloaded the preprocessed data and pretrained weights (Section 2.2), you can skip Steps 1 and 2 and go directly to Step 3: Evaluation.

Step 1 — Extract SDF Samples

Configuration: config_files/extract_sdf_ShapeNetAD.yaml

python data/extract_sdf_ShapeNetAD.py

Step 2 — Train SDF Model

Configuration: config_files/train_sdf_ShapeNetAD.yaml

python Train/train_sdf_ShapeNetAD.py

Step 3 — Evaluate

Update the dataset path in config_files/test_ShapeNetAD.yaml:

dataset:
  name: ShapeNetAD
  dataset_dir: /path/to/Anomaly-ShapeNet-v2/dataset/16384   # ← modify here

Run:

python Test/AD_test.py --config config_files/test_ShapeNetAD.yaml

Implementation Notes:

• The Chamfer distance in the PAM module was replaced with a directed variant for Real3D-AD, differing slightly from the paper but yielding better stability and efficiency.

• Adjusting the voxel_size parameter improves registration accuracy. The optimal per-class values (used for the best results reported in the paper) are provided in config_files/voxel_sizes.yaml, with a default of 0.03.

🧩 Reconstruction

1. Modify the YAML Configuration File

Update the checkpoint_path in config_files/reconstruct_mesh_ShapeNetAD.yaml:

checkpoint_path: results/ShapeNetAD/runs_sdf/ # default path 
mesh_save_dir: results/ShapeNetAD/reconstruct_mesh/

2. Run the Reconstruction Script

After modifying the YAML configuration file, run the reconstruction script using the following command:

#for Anomaly-ShapeNet dataset
python scripts/reconstruct_mesh.py config_files/reconstruct_mesh_ShapeNetAD.yaml
#for Real3D-AD dataset
python scripts/reconstruct_mesh.py config_files/reconstruct_mesh_Real3D_AD.yaml

Citation

If you find PASDF useful in your research, please cite:

@article{zheng2025bridging,
  title={Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation},
  author={Zheng, Bozhong and Gan, Jinye and Xu, Xiaohao and Li, Wenqiao and Huang, Xiaonan and Ni, Na and Wu, Yingna},
  journal={arXiv preprint arXiv:2505.24431},
  year={2025}
}