LGASS — An improved Graph Attention Network for Semantic Segmentation of Industrial Point Clouds in Automotive Battery Sealing Nail Defect Detection

Note: the codebase was renamed from LGANet → LGASS to match the published method name. The paper title remains as above for searchability. The old GitHub URL still redirects.

News

• EAAI 2026 — paper accepted (volume 163, pages 112793).
• Dataset released on HuggingFace as a gated repo: https://huggingface.co/datasets/vpan1226/OPT-SND.

Defect Classes

The dataset stores 8 raw classes on disk; the paper reports a merged 6-class evaluation (Background1 ∪ Background2 → Normal, Burst ∪ Burst2 → Burst).

Reported (paper, 6 cls)	Raw (on disk, 8 cls)
Normal	Background1 (0), Background2 (5)
Burst	Burst (1), Burst2 (6)
Pit	Pit (2)
Stain	Stain (3)
Warpage	Warpage (4)
Pinhole	Pinhole (7)

Data Acquisition

The dataset used in this project is collected from real-world industrial production lines for automotive battery sealing nail inspection.

All point cloud data are acquired using industrial 3D vision systems deployed in controlled manufacturing environments, covering various defect types such as surface deformation, misalignment, and structural anomalies of sealing nails. The image acquisition and imaging module consists of a Laser Profiler as a 3D scanning system (OPT-LPC20 Laser Profiler) with a mobile robotic platform and a test specimen. The laser profiler is configured in linear scanning mode. The system uses a blue laser stripe (wavelength 405nm) with a profile acquisition rate of 10,000 profiles per second, ensuring high fidelity on metallic reflective surfaces. During scanning, the sealing nails were placed on a precision rotary stage to ensure uniform coverage. The scanning speed was set to 10mm/s, and each profile captured 3000 points across the horizontal axis, resulting in a lateral resolution of ~0.05mm and vertical depth accuracy of <0.01mm. All scans were manually verified for completeness and noise artifacts.The battery is mounted on the mobile robot platform, and the calibrated 3D line laser scanning camera projects laser stripes onto the sealing nail area of the battery. Through synchronized image data acquisition during planar scanning trajectories, surface data of thesealing nail specimen are obtained.The data are manually annotated following a strict quality control process to ensure labeling accuracy and consistency.

Architecture

LGASS is built on the PyTorch deep-learning framework with two novel modules on top of a graph-attention encoder–decoder:

• LGAF (Local Graph Attention Filter) — exploits normal-difference attention to detect minute surface irregularities.
• SAG-Pooling (Spatial Attention Graph Pooling) — preserves geometry-critical points during downsampling.

See model/sem/GraphAttention.py and model/sem/network.py.

Results

Comparison on the sealing-nail dataset (paper Table 3, bold = best, underline = 2nd best):

Method	OA	mAcc	mIoU	Burst	Pit	Stain	Warpage	Pinhole
PointNet	83.29	72.83	46.21	32.27	34.96	19.97	65.37	36.04
PointNet++	89.33	84.65	62.67	47.91	59.32	42.37	81.64	46.53
GACNet	96.05	87.84	69.41	56.02	67.83	54.54	86.96	59.94
PTv1	98.77	90.96	72.83	68.54	62.83	58.37	89.67	58.37
OA-CNN	98.72	91.13	74.04	64.89	66.91	64.87	88.37	60.04
PTv3	99.03	94.96	75.13	77.32	67.98	60.44	89.17	56.21
LGASS	99.47	92.37	79.23	76.22	72.17	71.33	91.61	64.95

Note on reproduction: the numbers above are the published values from paper Table 3. A locally re-trained checkpoint matching these numbers is not yet bundled with this code release — the original checkpoint is no longer on hand and a 300-epoch retrain is in progress. To reproduce yourself, run bash scripts/reproduce_paper.sh (uses configs/paper.yaml); expect drift of ±0.5% on the headline metrics due to non-determinism in pointops kernels and CUDA / PyTorch version differences. Plan for several days of GPU time per the schedule in Table 2.

Ablation

Table 4 — module ablation (mIoU):

ID	LGAF	SAG-Pooling	OA	mIoU
I	✗	✗	96.92	72.93
II	✓	✗	97.38	77.17
III	✗	✓	97.96	74.81
IV	✓	✓	99.47	79.23

Table 5 — input-type ablation:

Input	OA	mAcc	mIoU
Coordinates only	99.42	90.92	76.94
Coordinates + Normals	99.51	91.43	78.31
Normals only (default)	99.47	92.37	79.23

Hyperparameters

configs/paper.yaml is the default and follows paper Table 2:

Parameter	Value
Epoch	300
Batch size	8
Number of points	16384
Optimizer	AdamW
Learning rate	0.01
LR decay	Multi-step

configs/original.yaml preserves the values originally hard-coded in train.py (AdamW, lr=0.001, batch=4, StepLR every 20 epochs by 0.5x) for transparency. We cannot confirm which set produced the paper's headline numbers without retraining both.

Dataset Structure

After downloading from HuggingFace (gated; see https://huggingface.co/datasets/vpan1226/OPT-SND):

data/sealingNail_npz/
├── train/   *.npz
└── test/    *.npz

Class distribution (paper Table 1, merged 6-class view):

Class	Count
Normal	271
Burst	170
Pit	112
Stain	102
Warpage	132
Pinhole	142

Each .npz contains a single key points, shape (N, 7), columns [x, y, z, nx, ny, nz, label].

Installation

Recommended: Use Conda

1. Create a new Conda environment:

   conda env create -f environment.yml

2. Activate the environment:

   conda activate sealingnail

3. Install PyTorch with CUDA:

   conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

4. Install other dependencies:

   conda install numpy scipy scikit-learn
   conda install -c conda-forge open3d
   conda install tqdm tensorboard h5py

5. Ensure C++/build tools (Linux):

   sudo apt install g++
   sudo apt install build-essential
   sudo apt install ninja-build
   ninja --version

6. Install custom pointops library:

   cd lib/pointops
   python setup.py install

7. Verify installation:

   python -c "import torch; print(torch.__version__); print(torch.cuda.is_available())"

Usage

Train:

python train.py --config configs/paper.yaml
# or use the original train.py values:
python train.py --config configs/original.yaml

Inference (single sample):

python demo.py --input data/test_sample.ply --model checkpoints/best_model.pth --output results/

Batch inference:

python batch_demo.py --input data/test_folder/ --model checkpoints/best_model.pth --output results/

Visualize a sample:

python visualize.py --input data/sealingNail_npz/test/sample_001.npz --save out.png --no-window

Generate a per-batch HTML report:

python export_report.py --input results/ --output reports/

Phase 2 (after retraining):

python evaluate.py --checkpoint logs/paper_run/best.pth --split test --output results/eval.md
bash scripts/reproduce_paper.sh
python scripts/upload_to_hf.py --checkpoint logs/paper_run/best.pth --repo vpan1226/LGASS

Citation

If you consider our work useful, please cite:

@article{pan2026improved,
  title   = {An improved graph attention network for semantic segmentation of industrial point clouds in automotive battery sealing nail defect detection},
  author  = {Pan, Wei and Wu, Yuhao and Tang, Wenming and Lu, Qinghua and Zhang, Yunzhi},
  journal = {Engineering Applications of Artificial Intelligence},
  volume  = {163},
  pages   = {112793},
  year    = {2026},
  publisher = {Elsevier}
}

License

Apache 2.0. See LICENSE.

Acknowledgements

This work was supported in part by OPT Machine Vision through internal research and development projects on industrial 3D vision and point cloud inspection systems.

The authors would like to thank all collaborators involved in data collection, annotation, and system deployment for automotive battery sealing nail inspection. We also appreciate the constructive feedback from anonymous reviewers, which helped improve the quality and clarity of this work.

In addition, we acknowledge the open-source contributions from the community. This project benefits from several excellent open-source works in point cloud processing and graph-based learning, including but not limited to:

• PointNet++
• DGCNN
• DeepGCN
• PointNeXt
• PointOps

We sincerely thank the authors of these works for making their code publicly available, which greatly facilitated reproducibility and comparative evaluation. Parts of the implementation are adapted or inspired by prior works, with appropriate modifications for industrial point cloud segmentation scenarios.