BGSPCD-v4-robust: Robust Benchmark Geometric Primitive Point Cloud Dataset

1. Overview

BGSPCD-v4-robust is a standardized point cloud dataset designed for evaluating 3D geometric primitive classification, topology inference, and bounded surface fitting in unstructured robotic grasp planning.

The dataset explicitly evaluates algorithm robustness under simulated single-view observation, optical sensor noise, partial truncation, spatial occlusion, sparse sampling, and arbitrary $SE(3)$ pose transformations.

2. Geometric Primitives

All samples belong to three canonical topological classes (strictly 1-to-1 mapped with neural classification logits):

• Class 0 (cuboid): Length, width, and height parameterized boxes.
• Class 1 (frustum): Bottom radius, top-to-bottom ratio, and axial height parameterized truncated cones and cylinders.
• Class 2 (ellipsoid): Tri-axial semi-axes parameterized ellipsoids.

3. Observation Perturbation Modes

Each shape family consists of 8 observation variants sharing identical canonical parameters and pose:

1. full_surface_prior: Uniform ground-truth surface sampling (32,768 points candidate, 2,048 sampled).
2. single_view: Standard synthetic ray-casting single-view depth camera observation.
3. single_view_clean: Single-view point cloud without sensor noise.
4. single_view_occluded: Realistic synthetic occlusion masking.
5. single_view_partial: High-ratio boundary truncation.
6. single_view_rotated: Arbitrary $SE(3)$ rotational perturbations.
7. single_view_sensor: Realistic depth camera Gaussian noise and lateral quantization.
8. single_view_sparse: Low-density subsampled surface points.

4. Archive Layout (Dual-Track)

bgspcd-v4-robust/
├── README.md                  # Dataset Card & documentation
├── dataset_metadata.json      # Camera intrinsics, primitive ranges & generator seed
├── manifest.jsonl             # Master sample index (sample_id, pose, dimensions, label)
├── samples.tar.gz             # Compressed core point cloud archives (.npz)
│   └── samples/
│       ├── train/ {cuboid, frustum, ellipsoid}
│       ├── val/   {cuboid, frustum, ellipsoid}
│       └── test/  {cuboid, frustum, ellipsoid}
└── legacy_cache.tar.gz        # Cache for PointNet/PointNet2 data loaders (.npz)

5. Quick Start (Python)

Download via Hugging Face Hub:

from huggingface_hub import hf_hub_download
import tarfile

# Download metadata and master index
manifest_path = hf_hub_download(repo_id="beta1scat/bgspcd-v4-robust", filename="manifest.jsonl", repo_type="dataset")
meta_path = hf_hub_download(repo_id="beta1scat/bgspcd-v4-robust", filename="dataset_metadata.json", repo_type="dataset")

# Download and extract the point clouds
samples_archive = hf_hub_download(repo_id="beta1scat/bgspcd-v4-robust", filename="samples.tar.gz", repo_type="dataset")
with tarfile.open(samples_archive, "r:gz") as tar:
    tar.extractall(path="./data/bgspcd_v4_robust")

Read a Sample:

import numpy as np

# Load sample .npz
data = np.load("./data/bgspcd_v4_robust/samples/test/cuboid/cuboid_000008_single_view.npz")
points = data["points"]  # (2048, 3) float32
label = int(data["label"])  # 0, 1, or 2

print(f"Loaded point cloud shape: {points.shape}, label: {label}")