3D point-cloud semantic-segmentation dataset for industrial defect detection on automotive battery sealing nails. Released alongside the EAAI 2026 paper "An improved Graph Attention Network for Semantic Segmentation of Industrial Point Clouds in Automotive Battery Sealing Nail Defect Detection."

Code: https://github.com/WillPANSUTD/LGASS · Paper: EAAI 2026, vol. 163, pp. 112793.

Gated access

This dataset is gated. To download, complete all three steps:

Authenticate locally:
```
huggingface-cli login
```
Provide a token from https://huggingface.co/settings/tokens.
Accept the terms: open https://huggingface.co/datasets/vpan1226/OPT-SND in a browser, click Access repository, fill in the form, agree to the conditions.
Wait for approval (manual review).

Without all three steps, load_dataset(...) and hf_hub_download(...) will raise 401 / 403.

Quick start

from datasets import load_dataset

# Requires gated-access steps above.
ds = load_dataset("vpan1226/OPT-SND")
sample = ds["train"][0]
points = sample["points"]   # (P, 7) float32 — see Data fields below

Or load a single file directly:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="vpan1226/OPT-SND",
    repo_type="dataset",
    filename="sealingNail_npz/train/sample_0001.npz",
)
arr = np.load(path, allow_pickle=True)["points"]  # (P, 7) float32
xyz = arr[:, 0:3]; feats = arr[:, 3:6]; labels = arr[:, 6].astype("int64")

Dataset summary

OPT-SND contains industrial 3D scans of automotive battery sealing nails, captured by an OPT-LPC20 line laser profiler in a controlled production environment. Each scan is a per-point-cloud sample with surface coordinates, normals, and a per-point semantic label drawn from an 8-class raw schema (described below). The dataset is intended for benchmarking semantic-segmentation methods on highly reflective metallic surfaces where color/intensity cues are unreliable.

Dataset structure

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

Sample counts and split sizes are visible to authenticated viewers on the HF web UI; this card avoids hard-coding them so it stays accurate as the dataset evolves.

Data fields

Each .npz contains a single key points, shape (N, 7), dtype float32:

Column	Field	Description
0	x	X coordinate (mm)
1	y	Y coordinate (mm)
2	z	Z coordinate (mm)
3	nx	surface-feature / normal x
4	ny	surface-feature / normal y
5	nz	surface-feature / normal z
6	label	semantic class id (raw 8-class, 0..7)

Cast column 6 to int64 after loading.

Class definitions

Raw schema (8 classes, as stored on disk):

ID	Name	Description
0	Background1	Background region / nominal surface (A)
1	Burst	Crack / rupture defect
2	Pit	Concave depression
3	Stain	Surface discoloration / contamination
4	Warpage	Out-of-plane deformation
5	Background2	Background region / nominal surface (B)
6	Burst2	Burst variant
7	Pinhole	Small puncture

Paper evaluation view (6 classes, after merge):

The accompanying EAAI 2026 paper reports metrics on a merged 6-class taxonomy:

Normal = Background1 ∪ Background2
Burst = Burst ∪ Burst2
Pit, Stain, Warpage, Pinhole unchanged.

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

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

Acquisition setup

Camera: OPT-LPC20 line laser profiler, blue laser (405 nm).
Acquisition rate: 10,000 profiles / second.
Scan speed: 10 mm/s (mobile robotic platform).
Lateral resolution: ~0.05 mm.
Depth accuracy: < 0.01 mm.
Points per profile: 3,000 across the horizontal axis.

Annotation protocol

All scans were manually annotated by trained operators following a strict quality-control protocol; multi-annotator review was used on ambiguous samples to enforce labeling consistency.

Considerations

Class imbalance. Defect classes are far rarer than the background; loss reweighting is recommended (the LGASS reference implementation derives weights from class frequency).
Single-domain coverage. All scans come from one production line and one camera setup. Generalization to other lines may require fine-tuning or domain-adaptation.
No personal data. No faces, no identifying information; samples are mechanical parts only.

Citation

@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.

You need to agree to share your contact information to access this dataset

OPT-SND — Sealing Nail Defect Dataset