Datasets:
metadata
license: mit
task_categories:
- graph-ml
tags:
- synthetic
- geometry
- graph-neural-network
- 3d
- centerline
- pipe-structures
pretty_name: Shape Geometry Dataset
size_categories:
- 10K<n<100K
Shape Geometry Dataset
Synthetic graph-based centerline representations of 3D geometric motifs (pipe-like structures).
JSON Schema
dataset.json is an array of shape records. Each record:
{
"category": "arc_90",
"nodes": [[x, y, z], ...],
"edges": [[i, j], ...],
"features": {
"curvature": [0.0, 0.1, ...],
"segment_angle": [0.0, 160.5, ...]
}
}
| Field | Type | Description |
|---|---|---|
category |
string | Shape class label (e.g. straight, arc_90, corner) |
nodes |
float[][] (N×3) | 3D points with ~1-unit spacing between neighbors |
edges |
int[][] (E×2) | Index pairs connecting nodes |
features.curvature |
float[] (N) | Menger curvature per node (0 at endpoints/hubs) |
features.segment_angle |
float[] (N) | Angle in degrees between edges at each node |
Categories
| Category | What it is | Samples |
|---|---|---|
straight |
Random-length line segment (5–20 units) |  |
arc_10 |
10° gentle arc, R=10 |  |
arc_20 |
20° arc |  |
arc_30 |
30° arc |  |
arc_40 |
40° arc |  |
arc_50 |
50° arc |  |
arc_60 |
60° arc |  |
arc_70 |
70° arc |  |
arc_80 |
80° arc |  |
arc_90 |
90° quarter-circle bend |  |
arc_100 |
100° arc |  |
arc_110 |
110° arc |  |
arc_120 |
120° wide arc |  |
arc_130 |
130° arc |  |
arc_140 |
140° arc |  |
arc_150 |
150° near-semicircle |  |
arc_160 |
160° arc |  |
arc_170 |
170° near-full semicircle |  |
corner |
Two straight arms at a sharp 90° angle |  |
junction_T |
Through-pipe with perpendicular branch (T-shape) |  |
junction_Y |
Trunk splitting into two splayed arms (Y-shape) |  |
Why this structure?
- Nodes + edges = graph — directly loadable into GNN frameworks (PyTorch Geometric, DGL).
- Unit spacing — standardises graph density across shapes; a 15-unit straight and a 90° arc at R=10 both have proportional node counts.
- Random rigid pose — every sample gets a random 3D rotation + translation so the model can't memorise orientation.
- Per-node features — curvature and segment angle give the network local geometric cues beyond raw xyz, helping distinguish arcs from straights even when node counts overlap.
- Junctions share a hub node — T and Y shapes have one node connected to 3 edges, matching real pipe topology and giving the GNN a clear topological signal.
Regenerate
python generate_geometry_dataset.py
Generates 10000 samples per category (17 arc classes + straight + corner + junction_T + junction_Y = 21 categories, 21k samples total).