sdf-sampler 0.6.0

Auto-analysis and sampling of point clouds for SDF (Signed Distance Field) training data generation

sdf-sampler

Auto-analysis and sampling of point clouds for SDF (Signed Distance Field) training data generation.

A lightweight, standalone Python package for generating SDF training hints from point clouds. Automatically detects SOLID (inside) and EMPTY (outside) regions and generates training samples suitable for SDF regression models.

Installation

pip install sdf-sampler

For additional I/O format support (PLY, LAS/LAZ):

pip install sdf-sampler[io]

Command-Line Interface

sdf-sampler provides a CLI for common workflows:

# Run as module
python -m sdf_sampler --help

# Or use the installed command
sdf-sampler --help

Commands

pipeline - Full workflow (recommended)

Run the complete pipeline: analyze point cloud → generate samples → export.

# Basic usage
sdf-sampler pipeline scan.ply -o training_data.parquet

# With options
sdf-sampler pipeline scan.ply \
    -o training_data.parquet \
    -n 50000 \
    -s inverse_square \
    --save-constraints constraints.json \
    -v

Options:

• -o, --output: Output parquet file (default: <input>_samples.parquet)
• -n, --total-samples: Number of samples to generate (default: 10000)
• -s, --strategy: Sampling strategy: constant, density, inverse_square (default: inverse_square)
• -a, --algorithms: Specific algorithms to run (default: all)
• --save-constraints: Also save constraints to JSON
• --seed: Random seed for reproducibility
• -v, --verbose: Verbose output

analyze - Detect regions

Analyze a point cloud to detect SOLID/EMPTY regions.

sdf-sampler analyze scan.ply -o constraints.json -v

Options:

• -o, --output: Output JSON file (default: <input>_constraints.json)
• -a, --algorithms: Algorithms to run (see below)
• --no-hull-filter: Disable hull filtering
• -v, --verbose: Verbose output

sample - Generate training samples

Generate training samples from a constraints file.

sdf-sampler sample scan.ply constraints.json -o samples.parquet -n 50000

Options:

• -o, --output: Output parquet file
• -n, --total-samples: Number of samples (default: 10000)
• -s, --strategy: Sampling strategy (default: inverse_square)
• --seed: Random seed
• -v, --verbose: Verbose output

info - Inspect files

Show information about point clouds, constraints, or sample files.

sdf-sampler info scan.ply
sdf-sampler info constraints.json
sdf-sampler info samples.parquet

Python SDK

Quick Start

from sdf_sampler import SDFAnalyzer, SDFSampler, load_point_cloud

# 1. Load point cloud (supports PLY, LAS, CSV, NPZ, Parquet)
xyz, normals = load_point_cloud("scan.ply")

# 2. Auto-analyze to detect EMPTY/SOLID regions
analyzer = SDFAnalyzer()
result = analyzer.analyze(xyz=xyz, normals=normals)
print(f"Generated {len(result.constraints)} constraints")

# 3. Generate training samples
sampler = SDFSampler()
samples = sampler.generate(
    xyz=xyz,
    constraints=result.constraints,
    strategy="inverse_square",
    total_samples=50000,
)

# 4. Export to parquet
sampler.export_parquet(samples, "training_data.parquet")

SDFAnalyzer

Analyzes point clouds to detect SOLID and EMPTY regions.

from sdf_sampler import SDFAnalyzer
from sdf_sampler.config import AnalyzerConfig, AutoAnalysisOptions

# With default config
analyzer = SDFAnalyzer()

# With custom config
analyzer = SDFAnalyzer(config=AnalyzerConfig(
    min_gap_size=0.10,      # Minimum gap for flood fill
    max_grid_dim=200,       # Maximum voxel grid dimension
    cone_angle=15.0,        # Ray propagation cone angle
    hull_filter_enabled=True,  # Filter outside X-Y hull
))

# Run analysis
result = analyzer.analyze(
    xyz=xyz,                    # (N, 3) point positions
    normals=normals,            # (N, 3) point normals (optional)
    algorithms=["flood_fill", "voxel_regions"],  # Which algorithms to run
)

# Access results
print(f"Total constraints: {result.summary.total_constraints}")
print(f"SOLID: {result.summary.solid_constraints}")
print(f"EMPTY: {result.summary.empty_constraints}")

# Get constraint dicts for sampling
constraints = result.constraints

Analysis Algorithms

Algorithm	Description	Output
flood_fill	Detects EMPTY (outside) regions by ray propagation from sky	Box or SamplePoint constraints
voxel_regions	Detects SOLID (underground) regions	Box or SamplePoint constraints
normal_offset	Generates paired SOLID/EMPTY boxes along surface normals	Box constraints
normal_idw	Inverse distance weighted sampling along normals	SamplePoint constraints
pocket	Detects interior cavities	Pocket constraints

SDFSampler

Generates training samples from constraints.

from sdf_sampler import SDFSampler
from sdf_sampler.config import SamplerConfig

# With default config
sampler = SDFSampler()

# With custom config
sampler = SDFSampler(config=SamplerConfig(
    total_samples=10000,
    inverse_square_base_samples=100,
    inverse_square_falloff=2.0,
    near_band=0.02,
))

# Generate samples
samples = sampler.generate(
    xyz=xyz,                     # Point cloud for distance computation
    constraints=constraints,      # From analyzer.analyze().constraints
    strategy="inverse_square",    # Sampling strategy
    seed=42,                      # For reproducibility
)

# Export
sampler.export_parquet(samples, "output.parquet")

# Or get DataFrame
df = sampler.to_dataframe(samples)

Sampling Strategies

Strategy	Description
constant	Fixed number of samples per constraint
density	Samples proportional to constraint volume
inverse_square	More samples near surface, fewer far away (recommended)

Constraint Types

The analyzer generates various constraint types:

• BoxConstraint: Axis-aligned bounding box
• SphereConstraint: Spherical region
• SamplePointConstraint: Direct point with signed distance
• PocketConstraint: Detected cavity region

Each constraint has:

• sign: "solid" (negative SDF) or "empty" (positive SDF)
• weight: Sample weight (default 1.0)

I/O Helpers

from sdf_sampler import load_point_cloud, export_parquet

# Load various formats
xyz, normals = load_point_cloud("scan.ply")    # PLY (requires trimesh)
xyz, normals = load_point_cloud("scan.las")    # LAS/LAZ (requires laspy)
xyz, normals = load_point_cloud("scan.csv")    # CSV with x,y,z columns
xyz, normals = load_point_cloud("scan.npz")    # NumPy archive
xyz, normals = load_point_cloud("scan.parquet") # Parquet

# Export samples
export_parquet(samples, "output.parquet")

Output Format

The exported parquet file contains columns:

Column	Type	Description
x, y, z	float	3D position
phi	float	Signed distance (negative=solid, positive=empty)
nx, ny, nz	float	Normal vector (if available)
weight	float	Sample weight
source	string	Sample origin (e.g., "box_solid", "flood_fill_empty")
is_surface	bool	Whether sample is on surface
is_free	bool	Whether sample is in free space (EMPTY)

Configuration Reference

AnalyzerConfig

Option	Default	Description
min_gap_size	0.10	Minimum gap size for flood fill (meters)
max_grid_dim	200	Maximum voxel grid dimension
cone_angle	15.0	Ray propagation cone half-angle (degrees)
normal_offset_pairs	40	Number of box pairs for normal_offset
idw_sample_count	1000	Total IDW samples
idw_max_distance	0.5	Maximum IDW distance (meters)
hull_filter_enabled	True	Filter outside X-Y alpha shape
hull_alpha	1.0	Alpha shape parameter

SamplerConfig

Option	Default	Description
total_samples	10000	Default total samples
samples_per_primitive	100	Samples per constraint (CONSTANT)
samples_per_cubic_meter	10000	Sample density (DENSITY)
inverse_square_base_samples	100	Base samples (INVERSE_SQUARE)
inverse_square_falloff	2.0	Falloff exponent
near_band	0.02	Near-band width
seed	0	Random seed

Integration with Ubik

sdf-sampler is the core analysis engine for Ubik, an interactive web application for SDF labeling. Use sdf-sampler directly for:

• Automated batch processing pipelines
• Integration into ML training workflows
• Custom analysis scripts

License

MIT