DeepInspect3D - 3D Mesh Inspection with Open3D-Based Point Cloud Alignment, Defect Localization, and Metric Evaluation

DeepInspect3D delivers an extensible 3D inspection pipeline for high-precision geometric defect detection in mechanical components. It integrates photorealistic synthetic rendering using Blender, point cloud generation, multi-stage filtering, FPFH+ICP alignment, and deviation field analysis against CAD ground truth, enabling automated metrology-grade evaluation and quality assurance.

---

Project Structure Overview

DeepInspect3D/
├── notebook/
│   └── inspection_pipeline.ipynb      # Unified inspection notebook
├── scripts/                          # Modular inspection stages
│   ├── render.py                     # Blender-based RGB/Depth/Normal rendering
│   ├── pointcloud.py                 # STL → PLY conversion
│   ├── preprocess_pointcloud.py      # Denoising, voxelization, normals
│   ├── register_pointcloud.py        # Global (FPFH) + ICP registration
│   ├── detect_defects.py             # Deviation analysis + defect mask
│   └── evaluation_metrics.py         # Chamfer, deviation, precision/recall
├── data/
│   ├── reference_master.stl          # CAD ground truth model
│   ├── test_scanned.stl              # QA scanned model
│   └── processed/
│       ├── reference_clean.ply
│       ├── test_clean.ply
│       └── test_aligned.ply
├── defects/
│   ├── deviation_heatmap.ply         # Color-mapped geometric deviation
│   └── binary_defect_mask.ply        # Thresholded defective regions
├── reports/
│   ├── metrics.json                  # Raw inspection metrics
│   └── inspection_report.json        # Final summary
├── blender_assets/                   # HDRI, textures, config
│   ├── textures/
│   ├── hdri/
│   └── config.yaml
├── requirements.txt                  # Python dependencies
└── README.md                         # Documentation and usage

---

Key Features

• CAD-to-Scan Inspection
  Compare scanned parts against their CAD references for precise geometry verification.

• Modular Point Cloud Pipeline
  Includes denoising, downsampling, normal estimation, global and local registration.

• Deviation Mapping + Masking
  Visual and binary extraction of out-of-tolerance regions using thresholded spatial error maps.

• Synthetic Data Support
  Procedural rendering of RGB, depth, and normal maps from .stl with Blender (HDRI + PBR materials).

• Quantitative Evaluation
  Metrics include:

  • Mean & Max Surface Deviation
  • Chamfer Distance (bidirectional)
  • Precision / Recall / F1 (if ground truth available)

---

📊 Outputs

Artifact	Path
Aligned Point Cloud	data/processed/test_aligned.ply
Deviation Heatmap	defects/deviation_heatmap.ply
Binary Defect Mask	defects/binary_defect_mask.ply
Metric Summary	reports/metrics.json
Final Inspection Report	reports/inspection_report.json

---

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Place STL Models

Put your reference and test .stl files in the data/ directory:

3. Launch Pipeline

jupyter notebook notebook/inspection_pipeline.ipynb

Follow the notebook to:

1. Convert STL to PLY
2. Clean and align clouds
3. Compute deviations
4. Save metrics and masks

---

🧪 Synthetic Data Rendering

Render realistic training/inference data using Blender’s headless mode:

cd scripts
blender -b -P render.py

Outputs are saved in blender_assets/outputs/:

• RGB images
• 16-bit depth maps
• Normal maps
• Camera intrinsics (JSON)

---

Requirements

Install via:

pip install -r requirements.txt

---

Applications

• Mechanical QA automation
• CAD-based geometry validation
• Metrology in manufacturing
• Research in 3D defect detection
• Synthetic dataset generation for 3D vision models

---

What's Next?

• Add Mesh-to-Mesh Deviation (Not Just Point Clouds)
• Surface Defect Segmentation using Deep Learning
• Integration with Industrial 3D Scanners

---

📩 Contact

I’m excited to connect and collaborate!

• Email: gbrohiith@gmail.com
• LinkedIn: https://www.linkedin.com/in/rohiithgb/
• GitHub: https://github.com/GBR-RL/

---

📚 License

This project is open-source and available under the MIT License.

---

🌟 If you like this project, please give it a star! 🌟