tryforge/pni-3d-benchmark

PNI-3D: Multimodal CAD & Geometric AI Benchmark

PNI-3D is a benchmark dataset designed for testing fine-grained 3D CAD matching, multimodal embedding alignment, and geometric retrieval under real-world sensor distortions.

---

Key Dataset Metrics

• 100 Parametric CAD Families: Mechanical engineering primitives (flanges, mounting brackets, U-channels, spur gears, shafts, heat sinks, valve housings, and more).
• 2,400 Total Assets: Pairs of nominal parts and functional variants, each in 12 representations.
• 12 Multimodal Representations per Part: Native parametric CAD, polygon meshes, point clouds, voxel grids, and 2D depth projections.
• Explicit Ground Truth: same-part groups, functional-variant groups, near-duplicate sets, negative pairs, and quantitative nuisance parameters.

---

12 Representations Per Part

Each parametric CAD instance is represented across 12 distinct formats to benchmark cross-modal embedding models:

Category	Representation	Description
Native CAD	_STEP	Boundary Representation (B-Rep) STEP file.
	_OBJ	Polygon mesh export.
Meshes	_MESH_FINE	High-density watertight triangulation.
	_MESH_COARSE	Decimated lightweight geometry (25% face count).
	_MESH_CORRUPTED	Non-manifold mesh with missing facets and flipped normals.
Point Clouds	_PCD_CLEAN	Dense uniform surface point sample (N = 10,000).
	_PCD_NOISY	Point cloud with Gaussian displacement along surface normals (σ = 0.5 mm).
	_PCD_SPARSE	Decimated low-density point cloud (N = 500).
Sensor / Spatial	_DEPTH_MAP	Multi-view 128 × 128 normalized depth array (.npy).
	_VOXEL	32 × 32 × 32 spatial occupancy grid (.npy).
Transformations	_NUISANCE_ROTATED	Arbitrary SO(3) rotation (Euler angles recorded per family).
	_NUISANCE_SCALED	Anisotropic scaling factors in [0.80×, 1.25×] (recorded per family).

---

Part Varieties & Parameter Splits

PNI-3D isolates nominal CAD geometry from functional variants to evaluate whether geometric models recognize subtle topological changes (like bolt-hole counts or thickness adjustments):

{
  "family_id": "FAM_001",
  "family_type": "flange",
  "nominal_parameters": {
    "outer_diameter": 83.74,
    "bore_diameter": 25.14,
    "thickness": 13.19,
    "bolt_holes": 8,
    "bolt_circle_radius": 32.33,
    "fillet_radius": 2.82
  },
  "variant_parameters": {
    "outer_diameter": 83.74,
    "bore_diameter": 25.14,
    "thickness": 18.19,
    "bolt_holes": 10,
    "bolt_circle_radius": 32.33,
    "fillet_radius": 0.0
  },
  "nuisance_parameters": {
    "rotation_angles_deg": [127.3, 34.677, -159.243],
    "scale_factors": [1.0313, 1.1479, 1.1917],
    "noise_sigma_mm": 0.5,
    "decimation_ratio": 0.25,
    "voxel_resolution": 32
  },
  "same_part_group": [
    "P_001_ORIGINAL_STEP",
    "P_001_ORIGINAL_MESH_FINE",
    "P_001_ORIGINAL_PCD_CLEAN",
    "P_001_ORIGINAL_VOXEL"
  ],
  "near_duplicates": [
    "P_001_ORIGINAL_MESH_FINE",
    "P_001_ORIGINAL_MESH_COARSE",
    "P_001_ORIGINAL_OBJ"
  ]
}

The 12 CAD categories are: flange, spur gear, mounting bracket, heat sink, valve housing, pipe fitting, hex bolt, stepped shaft, pulley, gusset, collar, and U-channel.

Repository Contents

Path	Description
metadata.parquet	100-row viewer table (schema above).
ground_truth_index.json	Full ground-truth index (all groups + nuisance parameters).
manifest.csv	Asset-level index (family, role, representation, path).
assets/FAM_*/	All 2,400 asset files (12 per part × 2 parts × 100 families).
generate_pni3d.py	Parametric generator (reproduce / scale to 600k).
evaluate_pni3d.py	Retrieval + Samples-to-Threshold benchmark harness.
pni3d_overview.png	Pipeline overview.

Quick Start Usage

from datasets import load_dataset

dataset = load_dataset("tryforge/pni-3d-benchmark")

# Inspect a sample part family
sample = dataset["train"][0]
print("Family Type:", sample["family_type"])
print("Nominal Bolt Holes:", sample["nominal_parameters"]["bolt_holes"])
print("Variant Bolt Holes:", sample["variant_parameters"]["bolt_holes"])

Contact & Support

If you encounter any issues accessing the dataset or require additional parameter sweeps, reach out directly:

Email: ravi@getforge.tech

LinkedIn: https://www.linkedin.com/in/raveekumar1/