Dataset Card for PnPCorrespondences

Ezharjan/PnPCorrespondences is a fully synthetic, image-free dataset of 2D-3D point correspondences with exact ground truth, built for benchmarking Perspective-n-Point (PnP) solvers, robust estimators, bundle adjustment and camera-calibration methods. Every sample is a set of 3D world points, their (noisy) 2D pixel observations and the exact camera intrinsics, lens distortion and extrinsics that produced them.

• Curated by: Aizierjiang Aiersilan
• License: cc-by-4.0
• Version: 1.0.0 (format 1.0)
• Size: 287,895 samples, 328,357,230 correspondences, 6.60 GB of HDF5
• Formats: HDF5 (arrays), Parquet and CSV (manifest), JSON (readable examples)

Dataset summary

Because there are no images, every source of error is controlled explicitly: Gaussian sub-pixel jitter, pixel quantization, outlier ratio (0 % to 95 %), outlier type (uniform replacement, swapped assignments or a mixture), lens model (pinhole, Brown-Conrady, Kannala-Brandt fisheye), field of view (5 deg telephoto to 175 deg fisheye) and scene structure (planar targets, room corners, volumes, mixed, depth-stratified corridors). The same geometry is observed under every noise condition, so any two conditions differ in exactly the factor that separates them and the effect of that factor is measurable without confounds.

scenes	400 (depth_stratified: 80, mixed: 80, planar_multi: 80, planar_single: 80, volumetric: 80)
camera views	19,193
samples (view x noise condition)	287,895
2D-3D correspondences	328,357,230
camera models	brown_conrady: 8689, kannala_brandt: 4624, pinhole: 5880
FOV classes	fisheye: 2648, narrow: 3809, normal: 7048, wide: 5688
HDF5 size	6.60 GB
master seed	20260902

Noise conditions

condition	samples
s0.00_q0_o0.00_uniform	19,193
s0.00_q1_o0.00_uniform	19,193
s0.10_q0_o0.00_uniform	19,193
s0.50_q0_o0.00_uniform	19,193
s0.50_q0_o0.05_uniform	19,193
s0.50_q0_o0.20_swap	19,193
s0.50_q0_o0.20_uniform	19,193
s0.50_q0_o0.50_swap	19,193
s0.50_q0_o0.50_uniform	19,193
s0.50_q0_o0.80_uniform	19,193
s0.50_q0_o0.95_uniform	19,193
s0.50_q1_o0.00_uniform	19,193
s1.00_q0_o0.00_uniform	19,193
s1.00_q1_o0.50_mixed	19,193
s2.00_q0_o0.00_uniform	19,193

Condition names read s<sigma>_q<quantized>_o<outlier ratio>_<outlier type>.

Uses

Direct use

• Benchmarking PnP and camera-resectioning solvers (DLT, P3P, AP3P, EPnP, IPPE, SQPnP, iterative refinement) against exact poses rather than against another estimator.
• Measuring the breakdown point of robust estimators (RANSAC, USAC, MAGSAC++) with a ground-truth outlier mask, which makes inlier precision and recall observable.
• Single-view and multi-view camera calibration, including fisheye rigs, with exact intrinsics and distortion coefficients to compare against.
• Training and evaluating learned pose estimators and outlier classifiers: splits are by scene, and the manifest carries every factor as a column for stratification.
• Ablations over noise, quantization, contamination, lens model, field of view, scene structure and number of correspondences, one factor at a time.

Out-of-scope use

• There are no images, so the dataset says nothing about feature detection, description or matching; correspondences are given, and their errors are prescribed.
• It is a controlled geometric benchmark, not a capture of the physical world: there is no photometric response, motion blur, rolling shutter or scene semantics.
• Real-world performance should be confirmed on real captures; what this dataset isolates is the estimator's behaviour under a known error model.

Dataset structure

Splits and configurations

The split column assigns every row to a split, and one Parquet file per split carries exactly those rows:

split	file	scenes	samples
train	manifest_train.parquet	320	230,295
val	manifest_val.parquet	40	28,800
test	manifest_test.parquet	40	28,800

Scenes, not samples, are assigned to a split, so no 3D structure and no camera is shared across the boundary. Two configurations are declared: default exposes the three splits above, and manifest is the complete table in one piece.

from datasets import load_dataset

test = load_dataset("Ezharjan/PnPCorrespondences", split="test")                 # one split of the manifest
everything = load_dataset("Ezharjan/PnPCorrespondences", "manifest", split="all")  # every row

Files

• manifest.parquet / manifest.csv - one row per sample with every scalar factor (scene type, split, camera model, FOV, intrinsics, distortion coefficients, noise parameters, number of visible points, ...) and the HDF5 location of the arrays.
• metadata/dataset_stats.json, metadata/config_used.yaml - statistics and the exact generator configuration.
• metadata/validation_report.json - the validator's own report for this build.
• examples/*.json - small human-readable samples, one per (scene type, camera model), in strict RFC 8259 JSON (non-finite values are written as null, never as bare NaN/Infinity literals).
• HDF5 shards (one per scene type and part):
• hdf5/depth_stratified_000.h5
• hdf5/depth_stratified_001.h5
• hdf5/depth_stratified_002.h5
• hdf5/depth_stratified_003.h5
• hdf5/mixed_000.h5
• hdf5/mixed_001.h5
• hdf5/mixed_002.h5
• hdf5/mixed_003.h5
• hdf5/planar_multi_000.h5
• hdf5/planar_multi_001.h5
• hdf5/planar_multi_002.h5
• hdf5/planar_multi_003.h5
• hdf5/planar_single_000.h5
• hdf5/planar_single_001.h5
• hdf5/planar_single_002.h5
• hdf5/planar_single_003.h5
• hdf5/volumetric_000.h5
• hdf5/volumetric_001.h5
• hdf5/volumetric_002.h5
• hdf5/volumetric_003.h5

HDF5 layout

/scene_XXXXX/                     attrs: scene_type, split, num_points, seed, ...
    points_3d        (N, 3) float64   world coordinates [m]
    point_labels     (N,)   int16     plane index of each point, -1 for volumetric points
    /camera_XXX/                  attrs: distortion_model, image_width, image_height, fov_class, hfov_deg, ...
        K                (3, 3) float64   intrinsics [[fx, s, cx], [0, fy, cy], [0, 0, 1]]
        dist_coeffs      (5,) or (4,)     (k1, k2, p1, p2, k3) Brown-Conrady / (k1..k4) Kannala-Brandt
        pose_Rt          (4, 4) float64   world -> camera:  X_c = R X_w + t
        camera_center    (3,)   float64
        points_2d_clean  (M, 2) float64   exact projections of the visible points
        point_indices    (M,)   int32     index into points_3d
        depths           (M,)   float64   z_c of the visible points
        /condition_XXX/           attrs: noise_sigma, quantize, outlier_ratio, outlier_type, num_outliers, ...
            points_2d    (M, 2) float64   noisy observations
            outlier_mask (M,)   bool      True where the observation does not belong to its 3D point

Loading

from huggingface_hub import snapshot_download
import h5py, pandas as pd

root = snapshot_download("Ezharjan/PnPCorrespondences", repo_type="dataset")
manifest = pd.read_parquet(f"{root}/manifest.parquet")
row = manifest[(manifest.split == "test") & (manifest.outlier_ratio == 0.2)].iloc[0]
with h5py.File(f"{root}/{row.file}", "r") as f:
    cond = f[row.h5_path]
    cam = cond.parent
    scene = cam.parent
    X = scene["points_3d"][()][cam["point_indices"][()]]   # (M, 3)
    uv = cond["points_2d"][()]                             # (M, 2) noisy observations
    K, dist, Rt = cam["K"][()], cam["dist_coeffs"][()], cam["pose_Rt"][()]
    outliers = cond["outlier_mask"][()]

snapshot_download accepts allow_patterns, so a single scene type can be fetched without the rest, for example allow_patterns=["manifest.parquet", "hdf5/planar_single_*"].

Dataset creation

The dataset was produced by the pnpcorr generator (MIT license); the methodology is summarised below. The exact configuration is stored in metadata/config_used.yaml, and every array is a deterministic function of the master seed (20260902), so the dataset can be regenerated bit-for-bit. Camera model conventions follow OpenCV (cv2.projectPoints / cv2.fisheye.projectPoints); across the whole sampled camera population the projections agree with those functions to better than 1e-11 px.

Notes on the design, relevant when interpreting the data:

• Distortion coefficients are sampled as the effective radial displacement at the image corner and converted to raw polynomial coefficients, so mild and strong mean the same thing at every focal length and resolution.
• Each distortion polynomial is restricted to its injective domain, bounded for Brown-Conrady by the first zero of the full 2-D Jacobian determinant, so every stored observation has a unique pre-image and can be undistorted exactly.
• Noise is applied as Gaussian jitter, then outlier contamination, then optional quantization, so every stored observation lies on the sensor grid when quantize is true.
• Observations are not clipped to the image, so the noise statistics are exact at the border.

Every build ships the validator's report: metadata/validation_report.json records the number of consistency checks run over the released files and the number that failed.

License

The dataset is released under cc-by-4.0. Copyright (c) 2026 Aizierjiang Aiersilan.

Citation

@misc{PnPCorrespondences,
  title        = {PnPCorrespondences: Synthetic 2D-3D Point Correspondences for Camera Calibration and PnP Benchmarking},
  author       = {Aizierjiang Aiersilan},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Ezharjan/PnPCorrespondences}
}