ViewGraphBench

Synthetic view graphs with exact ground truth for the global stage of Structure-from-Motion.

Toolkit and generator on GitHub | viewgraphforge on PyPI | Citation | MIT license

ViewGraphBench is a benchmark for rotation averaging, translation averaging, SE(3) pose-graph optimisation and the detection of outlier edges. A node of a view graph is a calibrated camera with a known pose. An edge is a measured relative pose between two cameras (rotation, unit translation direction and a metric translation for pose-graph optimisation) together with its standard deviations, the number of feature matches behind it, geometric covariates, its true error and a label that says whether and how it is corrupted.

Each graph is the output of a simulated global-SfM front end on a procedural 3-D scene. The cameras follow one of seven capture patterns and observe the scene points, the edges follow from co-visibility and an image-retrieval step, and the measurement errors depend on the number of matches and on the viewing geometry. Scenes with repeated structure (symmetric objects and monuments, periodic facades and corridors, twin buildings) produce wrong relative poses that agree with each other around cycles, which cycle-consistency checks cannot detect.

The graphs were generated with ViewGraphForge, an open-source toolkit (package viewgraphforge, command vgf) that also reads the files, runs the baseline solvers and the benchmark, and regenerates the dataset from its configuration.

quantity	value
graphs	45,495 in 16 splits
cameras	15,167,170 in total, 12 to 100,000 per graph
relative-pose measurements (edges)	290,885,926, 19.2 % of them labelled as outliers
camera-network families, scene types	7, 7
files	51,061: 45,495 graphs (.npz), 5,432 pose graphs (.g2o), tables and figures
size	102 GB
generator	viewgraphforge 1.0.0, global seed 20260909

One graph of each camera-network family (bench_topology, top views): cameras in blue, inlier edges in grey, random outliers in red, pink and purple, gross errors of the noise model in yellow.

Quick start

pip install -U "viewgraphforge[hub]"
hf download ezharjan/ViewGraphBench --repo-type dataset --local-dir ViewGraphBench --include "index.csv" --include "dataset_info.json" --include "splits/bench_topology/*"

import viewgraphforge as vgf
from viewgraphforge.benchmark import metrics, rotation_averaging, translation_averaging

g = vgf.load_graph("ViewGraphBench/splits/bench_topology/g0000_orbit_object_n300_r0.npz")
print(g.n_nodes, g.n_edges, g.is_outlier.mean())            # cameras, edges, outlier fraction
R = rotation_averaging.l1_irls(g)                            # global rotations from the relative ones
print(metrics.rotation_metrics(R, g.R_wc)["rot_median_deg"])  # median error in degrees
C = translation_averaging.lud(g, R)                          # camera centres up to a similarity
print(metrics.position_metrics(C, g.centers)["pos_median_rel"])

The full dataset is 102 GB. --include and --exclude select parts of it with shell-style patterns (* also matches /); with huggingface_hub 1.0 or newer, which the hub extra installs, both options can be repeated:

hf download ezharjan/ViewGraphBench --repo-type dataset --local-dir ViewGraphBench --include "splits/bench_*" --include "*.csv" --include "*.json" --include "manifest_sha256.txt"
hf download ezharjan/ViewGraphBench --repo-type dataset --local-dir ViewGraphBench --exclude "splits/train/*"
hf download ezharjan/ViewGraphBench --repo-type dataset --local-dir ViewGraphBench

The first command fetches the benchmark splits with the tables and the checksums, the second everything except the training split and the third the complete dataset.

index.csv lists every graph with its split, family, scene, size, noise and outlier settings, file path and checksum, so graphs can be selected before they are downloaded. The dataset viewer shows it, and datasets.load_dataset("ezharjan/ViewGraphBench", "index", split="graphs") loads it as a table.

Why synthetic view graphs

Global SfM back ends are usually evaluated on view graphs of real photo collections, whose reference poses are themselves reconstructions, or on synthetic graphs with random topology, isotropic noise and uniformly random outliers, which lack what makes real graphs difficult: weak edges with few matches and short baselines, long sequences with little overlap and outliers that agree with each other. This dataset provides

• exact ground truth for every camera, relative pose, measurement error and outlier, including the outlier type and, for repeated structure, the symmetry that caused it;
• measurements from a model of the front end: co-visibility from projection, occlusion and feature detection, retrieval-style edge selection, errors driven by the number of matches and the baseline-to-depth ratio, and optionally a simulated two-view estimator;
• sweeps that vary one factor at a time (size, noise, outlier ratio, outlier type, repeated structure, density, heavy tails, match threshold, scale noise), so that robustness can be reported as curves;
• a training split of 40,000 graphs, with validation and test splits from the same distribution, for learned outlier filters and graph neural networks;
• the toolkit that generated it, with loaders, baseline solvers, metrics, the benchmark protocol, validation and the figures of this page.

Splits

split	graphs	cameras per graph	design	varies
train	40,000	12 to 800	40,000 random draws, 30 to 800 cameras requested (log-uniform)	training; family, size, pixel noise, outlier ratio and type, retrieval depth, detection rate, noise model, heavy tails, scale noise and repeated structure drawn at random
val	2,000	18 to 799	2,000 random draws, 30 to 800 cameras requested (log-uniform)	validation, same distribution as train
test_wild	2,000	15 to 800	2,000 random draws, 30 to 800 cameras requested (log-uniform)	test, same distribution as train
test_wild_large	200	800 to 4,801	200 random draws, 800 to 5,000 cameras requested (log-uniform), analytic noise model only	test with larger graphs than in training
bench_scale	231	30 to 100,000	n_nodes in {50, 100, 250, 500, 1000, 2500, 5000, 10000, 25000, 50000, 100000} × 7 families × 3 repetitions	number of cameras
bench_noise	147	295 to 300	pixel_noise in {0.25, 0.5, 1, 2, 4, 8, 16} × 7 families × 3 repetitions; n_nodes 300	pixel noise of the image points
bench_outliers	168	293 to 300	outlier_ratio in {0, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6} × 7 families × 3 repetitions; n_nodes 300	fraction of random outliers
bench_outlier_mode	63	276 to 300	outlier_mode in {translation, rotation, both} × 7 families × 3 repetitions; n_nodes 300, outlier_ratio 0.3	part of the measurement a random outlier replaces
bench_symmetry	126	295 to 300	symmetry_confusability in {0, 0.1, 0.2, 0.35, 0.5, 0.75} × 7 cases with repeated structure × 3 repetitions; n_nodes 300, outlier_ratio 0	confusability of repeated structure
bench_density	126	262 to 300	retrieval_k in {3, 5, 10, 20, 40, 80} × 7 families × 3 repetitions; n_nodes 300	retrieval depth (views kept per camera)
bench_heavytail	84	296 to 300	heavy_tail_df in {0, 5, 3, 2} × 7 families × 3 repetitions; n_nodes 300, outlier_ratio 0	Student-t noise (degrees of freedom, 0 = Gaussian)
bench_matches	84	96 to 300	min_matches in {10, 20, 40, 80} × 7 families × 3 repetitions; n_nodes 300	minimum number of matches per edge
bench_topology	70	297 to 300	7 families × 10 repetitions; n_nodes 300	camera-network family at the default settings
bench_twoview	63	100 to 1,000	n_nodes in {100, 300, 1000} × 7 families × 3 repetitions; noise_model twoview	number of cameras, simulated two-view estimator
bench_twoview_symmetry	28	294 to 300	symmetry_confusability in {0.2, 0.5} × 7 cases with repeated structure × 2 repetitions; n_nodes 300, noise_model twoview, outlier_ratio 0	confusability of repeated structure, simulated two-view estimator
bench_pgo	105	497 to 500	scale_noise in {0.01, 0.03, 0.1, 0.2, 0.4} × 7 families × 3 repetitions; n_nodes 500, outlier_ratio 0.1	noise of the metric baseline lengths (SE(3) variant)

Settings that a split does not fix take the generator defaults: 1 px image noise, the analytic noise model, retrieval depth 20, at least 20 matches per edge, 10 % random outliers of mixed type (more often on edges with few matches), Gaussian noise, feature detection probability 0.8 and 5 % log-normal noise on the metric baseline lengths. In wild splits the sampler draws these settings for every graph, except the minimum number of matches. Only the largest connected component of a view graph is kept, so some graphs have fewer cameras than requested; the graph id records the requested number.

In the sweep splits, the seed of a graph depends on the case and the repetition but not on the level of the swept factor. All levels therefore share the scene, the cameras and the random numbers up to the step where the factor acts, and differences between levels are caused by the factor (in bench_scale and bench_twoview the factor is the number of cameras, so the cameras differ between levels, and the corridor, urban and terrain scenes also grow with it).

Statistics of each split (stats/splits.csv):

split	cameras	edges	mean degree	outlier fraction	symmetry-induced edges	inlier rotation error [deg]	inlier direction error [deg]
train	9,322,257	167,127,204	28.8	0.249	12,751,910	0.088	0.135
val	474,777	8,220,231	28.5	0.243	598,787	0.089	0.142
test_wild	483,442	8,602,980	29.4	0.248	633,759	0.088	0.134
test_wild_large	454,699	10,550,698	45.5	0.232	617,323	0.056	0.152
bench_scale	4,082,251	89,462,079	38.4	0.102	0	0.086	0.190
bench_noise	44,044	844,109	38.3	0.115	0	0.181	0.252
bench_outliers	50,288	965,432	38.4	0.269	0	0.083	0.110
bench_outlier_mode	18,825	362,601	38.5	0.301	0	0.076	0.107
bench_symmetry	37,728	892,434	47.3	0.149	138,512	0.074	0.112
bench_density	37,704	681,482	36.1	0.101	0	0.089	0.135
bench_heavytail	25,176	483,856	38.4	0.002	0	0.109	0.143
bench_matches	24,737	456,406	36.6	0.101	0	0.083	0.113
bench_topology	20,988	404,234	38.5	0.101	0	0.084	0.116
bench_twoview	29,394	585,337	36.2	0.103	0	0.076	0.143
bench_twoview_symmetry	8,380	204,173	48.8	0.183	36,456	0.074	0.094
bench_pgo	52,480	1,042,670	39.7	0.101	0	0.078	0.106

Cameras, edges and symmetry-induced edges are totals over the graphs of a split; degree and outlier fraction are means over its graphs; the errors are the median over its graphs of the median error of the inlier edges of each graph.

Composition of the dataset: families, splits, graph sizes, degrees, outlier fractions and inlier errors.

What a graph contains

Each graph is a compressed numpy archive (.npz) with the arrays below and a meta_json string that holds the generation parameters, the seed and summary statistics. A graph has n cameras and m edges; every camera pair is stored once, with i < j.

array	shape	type	content
node_q_wc	(n,4)	float64	world-to-camera rotation R_i as a unit quaternion
node_t_wc	(n,3)	float64	world-to-camera translation t_i [m], so that x_cam = R_i X + t_i
node_center	(n,3)	float64	camera centre C_i = -R_i^T t_i [m]
node_K	(n,4)	float64	intrinsics fx, fy, cx, cy [px]
node_image_size	(n,2)	int32	width, height [px]
node_seq_index	(n,)	int32	acquisition order (a random permutation for the unordered crowd family)
node_n_visible	(n,)	int32	scene points observed by the camera
node_gravity_cam	(n,3)	float64	world up direction in the camera frame, R_i [0,0,1]^T
edge_i, edge_j	(m,)	int32	camera indices, i < j
edge_q_meas	(m,4)	float64	measured relative rotation R_ij
edge_tdir_meas	(m,3)	float64	measured unit translation direction tdir_ij, in the frame of camera j
edge_t_meas	(m,3)	float64	measured metric relative translation t_ij [m] (SE(3) variant)
edge_q_true, edge_tdir_true, edge_t_true	(m,4), (m,3), (m,3)	float64	the same quantities computed from the ground-truth poses
edge_sigma_rot, edge_sigma_dir	(m,)	float32	reported noise level per axis of the rotation vector and per tangent axis of the direction [rad]: a standard deviation, or the scale of the Student-t distribution when heavy_tail_df > 0
edge_sigma_trans	(m,)	float32	approximate standard deviation of the metric translation per axis [m]
edge_n_matches, edge_n_covis	(m,)	int32	matches supporting the measured relative pose; co-visible scene points
edge_baseline, edge_median_depth	(m,)	float32	length of C_j - C_i; mean of the two cameras' median point depths [m]
edge_view_angle_deg	(m,)	float32	angle between the optical axes [deg]
edge_overlap_i, edge_overlap_j	(m,)	float32	fraction of each camera's points that the other camera also observes
edge_err_rot_deg, edge_err_dir_deg, edge_err_trans	(m,)	float32	true error of the measurement [deg, deg, m]
edge_outlier_type, edge_is_outlier	(m,)	int8, bool	outlier type (table below); edge_outlier_type != 0
edge_symmetry_id	(m,)	int32	row of the sym_* arrays for symmetry-induced outliers, otherwise -1
points_xyz, points_module, points_copy	(P,3), (P,), (P,)	float32, int32	scene points [m] (at most 20,000; not stored in bench_scale) and their repeated-structure labels (-1 for a unique point)
sym_R, sym_t, sym_module, sym_copy_from, sym_copy_to	(S,3,3), (S,3), (S,)	float64, int32	the S symmetry transforms X_to = R X_from + t behind the symmetry-induced outliers (only in graphs with such outliers)
obs_cam, obs_point, obs_xy	(K,), (K,), (K,2)	int32, float32	image observations of the stored points in normalised coordinates, with the pixel noise of the graph (graphs with stored points and up to 2,000 cameras, at most 300 per camera)

edge_outlier_type	name	measurement
0	none	the true relative pose with noise (inlier)
1	random_both	random rotation and random direction
2	random_rotation	random rotation; the direction is the noisy true one
3	random_translation	random direction; the rotation is the noisy true one
4	symmetry	the relative pose, with noise, to a virtual camera: the true camera moved by a symmetry of the scene, as when a matcher confuses two copies of a repeated structure; consistent around cycles (for the translational symmetries of periodic facades and corridors, the rotation is correct and only the direction is wrong)
5	gross_error	not injected: a measurement of the noise model whose error exceeds 10 deg in rotation or 20 deg in direction

Random rotations and directions are uniformly distributed. Because the true error of every edge is stored, other definitions of an outlier can be applied directly; the benchmark, for example, uses a rotation error above 5 deg (a direction error above 10 deg for the direction residual), so the outliers of translational symmetries count only for the direction residual.

Among the cases with repeated structure, the urban scene with twin buildings (manhattan family) produces few symmetry-induced edges.

Conventions: metres; world frame with +z up; OpenCV camera frame (x right, y down, z forward). Relative poses map camera i to camera j, R_ij = R_j R_i^T and t_ij = t_j - R_ij t_i (x_j = R_ij x_i + t_ij), so E_ij = [t_ij]_x R_ij satisfies x_j^T E_ij x_i = 0 and the baseline direction in the world frame is (C_j - C_i) / |C_j - C_i| = -R_j^T tdir_ij. Quaternions are stored scalar-last, (qx, qy, qz, qw), the order of SciPy and of the g2o format, with qw >= 0.

The .g2o files contain VERTEX_SE3:QUAT camera-to-world poses (R_i^T, C_i), initialised by chaining the measured relative poses along a maximum spanning tree weighted by the number of matches (not the ground truth), and EDGE_SE3:QUAT measurements Z_ij = T_i^-1 T_j = (R_ij^T, -R_ij^T t_ij) with the information matrix diag(I/sigma_trans^2, I/sigma_rot^2) in the order (x, y, z, qx, qy, qz). sigma_rot refers to the rotation vector, as GTSAM and SE-Sync read the block; g2o's own EdgeSE3 applies it to the vector part of the quaternion, about half the rotation vector, so multiply the rotation block by 4 for the same weighting in g2o. They exist for the graphs of up to 20,000 cameras of every split except train; vgf.write_g2o(graph, path) writes one for any graph.

One graph in detail: top and 3-D view, adjacency pattern and the measurement errors of each edge class.

Repeated structure (bench_symmetry): the orange edges are relative poses to a camera moved by a symmetry of the scene. They agree with each other around cycles.

Files

README.md, LICENSE
index.csv                      one row per graph: split, family, scene, size, settings, files, SHA-256
dataset_info.json              counts per split and the complete configuration
manifest_sha256.txt            SHA-256 of every .g2o and .npz file
splits/<split>/<graph_id>.npz  a graph (splits with up to 3,000 graphs)
splits/<split>/<graph_id>.g2o  the graph as an SE(3) pose graph, without the ground truth (see above)
splits/<split>/_split.json     design record: specification of the split, seed and parameters of every graph
splits/train/<a>-<b>/          graphs a to b of train, 1,000 per folder (00000-00999, 01000-01999, ...)
stats/                         graphs.csv (one row per graph), splits.csv, summary.json, summary.md
benchmarks/                    results.csv (one row per graph and method), leaderboard.csv/.md, sweeps.csv, run.json
figures/                       the figures of this page, galleries of every split, overviews of single graphs

Graph ids have the form g<index>_<family>_<scene>_n<cameras requested>[_<factor><value>]_r<repetition>, for example g0012_orbit_object_n300_conf0.5_r0, where the factor is a short code (px pixel noise, out outlier ratio, om outlier mode, conf confusability, k retrieval depth, df degrees of freedom, mm minimum matches, sc scale noise); sweeps over the number of cameras add no suffix. An id is unique within its split but can occur in several splits, so tables are joined on the pair (split, graph_id). The graphs of train come without .g2o files to keep the download small; vgf.write_g2o creates them from the .npz files.

Loading the data

With the toolkit, after downloading the benchmark splits:

import viewgraphforge as vgf

index = vgf.read_index("ViewGraphBench")                      # pandas DataFrame, one row per graph
rows = index[(index.split == "bench_outliers") & (index.outlier_ratio_target == 0.3)]
graphs = [vgf.load_graph(f"ViewGraphBench/{p}") for p in rows.npz_path]
g = graphs[0]
R_true, C_true = g.R_wc, g.centers        # ground truth: (n,3,3) rotations, (n,3) centres
R_ij, tdir_ij = g.R_meas, g.tdir_meas     # measurements (m,3,3), (m,3) of the edges g.edges (m,2)

With numpy only:

import json
import numpy as np

d = np.load("ViewGraphBench/splits/bench_topology/g0000_orbit_object_n300_r0.npz")
meta = json.loads(str(d["meta_json"]))    # parameters, seed, statistics
i, j, q = d["edge_i"], d["edge_j"], d["edge_q_meas"]

The .g2o files can be read by g2o and by GTSAM (readG2o) as they are. Their first line is a comment, which the reader of SE-Sync does not accept; remove it for SE-Sync.

How a graph is generated

Every graph is generated with a random generator seeded from its own seed:

1. Scene: a procedural point cloud with surface normals (object, facade, corridor, urban block, room, terrain or plaza). The corridor, urban, terrain and facade scenes grow with the number of cameras. Scenes with repeated structure contain identical copies of a module: the sectors of a symmetric object, facade columns, corridor segments, twin buildings or the faces of a monument.
2. Cameras: exactly n pinhole cameras (1600 x 1200 px) of one family: orbit (a ring around an object), dome (several rings on a hemisphere), corridor (forward motion along a street), lawnmower (an aerial survey), crowd (unordered photos of a landmark with varying focal lengths), walk (hand-held motion in a room) or manhattan (a vehicle on a street grid, with loop closures).
3. Visibility: a point is observed when it lies within the depth range, projects into the image, faces the camera (incidence angle below 80 deg), passes a z-buffer occlusion test and is detected (probability 0.8 by default); at most 3,000 points per image are kept.
4. Edges: in graphs of up to 400 cameras every camera pair is a candidate; larger graphs take the nearest neighbours in camera position and viewing direction, the three previous and next cameras in acquisition order and, for repeated structure, random pairs of cameras that see copies of the same module. Match counts are binomial in the number of co-visible points, with a probability that falls with the angle between the views and the scale change; repeated structure adds competing hypotheses, and the reported relative pose is drawn among those with at least min_matches matches, with a preference for the one with the most support. Each camera keeps the retrieval_k candidates with the most co-visible points (and up to cand_sym_k that look alike through repeated structure), and a candidate becomes an edge when its reported pose has at least min_matches matches. The largest connected component is kept.
5. Measurements: the reported relative pose plus noise, either from a closed-form model (analytic) or from a simulated two-view estimator (twoview). The analytic model is calibrated against the estimator: its errors scale with the pixel noise over the focal length and with 1/sqrt(matches), vary with the baseline-to-depth ratio rho as exp(b1 L + b2 L^2) with L = log(rho) and rho clipped to [0.08, 10] (below 0.08 the direction error also grows as 1/rho) and carry a log-normal per-edge factor, optionally with Student-t tails (heavy_tail_df, analytic model only); a direction that the baseline cannot determine is replaced by a random one. The estimator runs the normalised 8-point algorithm on the noisy correspondences, refines its solution and, separately, the reported pose (standing in for a minimal solver inside RANSAC) by Levenberg-Marquardt minimisation of the Sampson error, keeps the solution with the lower error and reports covariances from the Fisher information; an edge on which it fails keeps the analytic measurement.
6. Outliers: a fraction of the edges, chosen with probability inversely proportional to their number of matches, receives random rotations and/or directions; symmetry-induced outliers come from step 4. The metric translation for the SE(3) variant is the measured direction times the length of the reported relative translation with log-normal noise, and every error is recorded.

Baselines

vgf benchmark runs the baselines on the graphs of up to 3,000 cameras (pose-graph optimisation up to 1,500 cameras). Rotation errors are measured after a robust alignment of the global rotation. Positions are compared after a similarity alignment (translation averaging) or a rigid alignment (pose-graph optimisation) and divided by the diagonal of the bounding box of the true camera centres. Outlier detection is scored against the edges with a rotation error above 5 deg (a direction error above 10 deg for the direction residual). The tables give the mean over the graphs of a split of the per-graph median error or F1 score; graphs without labelled outliers are left out of the F1 means, and graphs on which a method failed are left out of its means.

This leaderboard covers 3,329 graphs of the splits test_wild, test_wild_large, bench_scale, bench_noise, bench_outliers, bench_outlier_mode, bench_symmetry, bench_density, bench_heavytail, bench_matches, bench_topology, bench_twoview, bench_twoview_symmetry and bench_pgo. The complete tables are in benchmarks/leaderboard.md and the per-graph results in benchmarks/results.csv.

Rotation averaging, median rotation error [deg] (best value of each split in bold):

split	spanning_tree	chordal_l2	chordal_l2+irls_cauchy	l1_irls	l1_irls_info
test_wild	31.9	16.3	9.94	7.63	8.88
test_wild_large	64.5	23.2	11.3	8.76	12.2
bench_scale	37.9	3.64	0.102	0.102	0.0541
bench_noise	41.2	2.44	0.216	0.216	0.106
bench_outliers	64.5	7.72	0.763	0.0921	6.42
bench_outlier_mode	68.6	11.4	0.106	0.105	2.26
bench_symmetry	0.271	20.9	22.3	5.92	5.15
bench_density	34	5.71	0.0932	0.0933	3.86
bench_heavytail	0.396	0.116	0.0929	0.0929	0.0297
bench_matches	35.5	2.83	0.0773	0.0773	1.21
bench_topology	32.9	2.71	0.0783	0.0783	0.0406
bench_twoview	28.8	2.86	0.0803	0.0803	0.0375
bench_twoview_symmetry	0.304	29	29.1	0.153	0.044
bench_pgo	25.5	2.51	0.0796	0.0796	0.0352

Translation averaging, median position error [% of the diagonal] (best value of each split in bold):

split	linear_cross@gt_rot	linear_cross@l1_irls	lud@gt_rot	lud@l1_irls
test_wild	16.9	16.9	6.2	6.41
test_wild_large	15.5	15.5	5.65	5.93
bench_scale	13.5	13.4	1.24	1.32
bench_noise	15	15	1.39	1.38
bench_outliers	13.8	13.8	2.32	2.37
bench_outlier_mode	12	12.2	2.32	2.38
bench_symmetry	14.3	14.2	6.21	7.86
bench_density	15.2	15.1	1.41	1.41
bench_heavytail	0.778	0.821	0.168	0.177
bench_matches	14.5	14.5	1.58	1.63
bench_topology	14.5	14.5	1.33	1.33
bench_twoview	14	14.1	1.26	1.26
bench_twoview_symmetry	17.9	17.6	7.33	7.72
bench_pgo	13.2	13.3	0.607	0.612

Pose-graph optimisation, median position error [% of the diagonal] (best value of each split in bold):

split	chain_init	lm_l2	lm_huber	lm_cauchy	lm_cauchy@rotavg_init
test_wild	7.62	12.7	5.51	6.23	3.76
test_wild_large	13.6	14.3	4.53	11.9	3.47
bench_scale	6.45	5.97	0.221	5.49	0.134
bench_noise	6.95	5.65	0.206	4.66	0.112
bench_outliers	12.5	11.5	0.864	11.6	0.114
bench_outlier_mode	14.4	13.2	0.59	12.7	0.113
bench_symmetry	0.649	15	10.5	0.107	4.23
bench_density	7.65	6.64	0.585	6.24	0.11
bench_heavytail	0.806	0.117	0.0987	0.0904	0.0901
bench_matches	7.86	6.11	0.544	6.51	0.424
bench_topology	6.29	6.15	0.187	5.02	0.0964
bench_twoview	5.74	5.86	0.191	4.45	0.104
bench_twoview_symmetry	0.704	18.4	15.3	0.124	6.68
bench_pgo	7.88	5.96	1.38	5.23	1.04

Outlier detection, F1 score (best value of each split in bold):

split	cycle_consistency	rot_residual@l1_irls	dir_residual@lud
test_wild	0.784	0.851	0.787
test_wild_large	0.775	0.849	0.768
bench_scale	0.997	0.999	0.902
bench_noise	0.950	0.991	0.905
bench_outliers	0.995	0.996	0.868
bench_outlier_mode	0.910	0.993	0.788
bench_symmetry	0.238	0.671	0.514
bench_density	0.978	0.997	0.883
bench_heavytail	0.813	0.945	0.740
bench_matches	0.996	0.999	0.910
bench_topology	0.998	0.999	0.908
bench_twoview	0.982	0.996	0.907
bench_twoview_symmetry	0.332	0.891	0.518
bench_pgo	0.997	0.999	0.921

Rotation averaging: spanning_tree chains the relative rotations along a maximum spanning tree weighted by the number of matches; chordal_l2 is the chordal least-squares solution (Martinec and Pajdla 2007), refined by Cauchy-weighted IRLS in chordal_l2+irls_cauchy; l1_irls is L1 averaging followed by IRLS (Chatterjee and Govindu 2013, 2018), with information weights in l1_irls_info.

Translation averaging: linear_cross (Govindu 2001) and lud (Özyeşil and Singer 2015) estimate the camera centres from the ground-truth rotations (@gt_rot) or from the l1_irls rotations (@l1_irls).

Pose-graph optimisation: chain_init is the spanning-tree chaining of the metric measurements alone; lm_l2, lm_huber and lm_cauchy are Levenberg-Marquardt optimisation with the named loss started from it, and lm_cauchy@rotavg_init starts from the l1_irls rotations and robustly solved centres (Carlone et al. 2015).

Outlier detection: cycle_consistency flags an edge when none of its triangles closes within 5 deg (at most 30 sampled triangles; edges without a triangle are kept), a simple form of the loop test of Zach et al. 2010; rot_residual@l1_irls and dir_residual@lud flag the measurements that disagree with the averaged rotations by more than 5 deg or with the averaged positions by more than 10 deg.

From left to right and top to bottom: rotation averaging against the fraction of random outliers; outlier detection against the confusability of repeated structure; rotation averaging against the pixel noise; rotation averaging against the number of cameras (up to 2,500). Each point is the mean over the graphs of one level (for the F1 score, over its graphs with labelled outliers).

Regenerating and extending the dataset

These commands regenerate the dataset and the files of this repository:

pip install viewgraphforge
vgf config full
vgf generate --config full --out data_full --workers 16
vgf validate --data data_full --workers 16
vgf stats --data data_full --workers 16
vgf benchmark --data data_full --splits "bench_*" "test_*" --workers 16
vgf visualize --data data_full
vgf export --data data_full --out data_hub
vgf card --data data_hub

vgf config full prints the configuration of this dataset, vgf generate continues where it stopped when it is run again after an interruption, and the last two commands produce the layout of this repository and this page.

The seed of every graph is derived from the global seed, the split and the place of the graph in the design (its index in a wild split, its case and repetition in a sweep or grid), so the result depends neither on the number of workers nor on the order of execution. With the same platform and library versions a regenerated graph is identical array by array; only the generation times in its metadata differ, so the file hashes differ between runs. Elsewhere the last bits of floating-point results can differ. That can flip a visibility or matching decision, after which the graph continues with a different random stream and becomes another sample of the same distribution. Regenerated on Linux with numpy 2.5.3 and scipy 1.18.1, 61 benchmark graphs of this copy (made on Windows) kept their edges, match counts and labels, while 3 of 7 training graphs (two with the twoview estimator, one with more than 400 cameras) did not; with numpy 2.4.4 and scipy 1.17.1, 26 of the 61 benchmark graphs differed as well. Regenerated copies are therefore compared with vgf validate and the statistics, not by file hash.

Generating the published copy took about 13 hours with 16 worker processes on a laptop with 16 logical CPUs and 32 GB of RAM. The 100,000-camera graphs of bench_scale take from 3 minutes to 2.7 hours each and up to 4.3 GB of memory; vgf generate schedules the graphs against a memory budget.

A configuration is a YAML file that lists the splits and their factor levels; vgf config smoke > my.yaml writes a small one to start from, and vgf generate --config my.yaml --out my_data builds a dataset in the same format.

Validation

vgf validate --data <folder> checks every graph: array shapes, index ranges and duplicate edges; unit quaternions and directions; the centres, gravity vectors, true relative poses, baselines and errors against the ground-truth poses and the measurements; that edge_is_outlier agrees with edge_outlier_type; the image observations against the poses and points; the .g2o file against the arrays (measurements, information matrices and initial poses); the row of the graph in index.csv; and the SHA-256 of its files listed in manifest_sha256.txt. Every file listed in the index or the manifest must exist, so run it on complete splits downloaded together with index.csv and manifest_sha256.txt.

Limitations

• There are no images. Feature detection and matching are statistical models, and appearance enters only through the confusability of repeated structure.
• The cameras are ideal pinhole cameras with known intrinsics, without lens distortion, rolling shutter or motion blur.
• The scenes are procedural layouts of seven types, not reconstructions of real places.
• The analytic noise model is calibrated against the simulated two-view estimator, not against real image pairs; the twoview graphs use the estimator itself, which sees only correct correspondences with Gaussian pixel noise (mismatched features and RANSAC are not simulated).
• Random outliers are uniform over rotations or directions, and symmetry-induced outliers follow the symmetry exactly; real matchers also produce partly wrong and mixed measurements.

Citation

If you use the dataset, please cite:

@misc{viewgraphbench2026,
  title        = {ViewGraphBench: Physically-Grounded Synthetic View Graphs for Global Structure-from-Motion Calibration},
  author       = {Aiersilan, Aizierjiang},
  year         = {2026},
  note         = {Dataset version 1.0.0, generated with ViewGraphForge},
  howpublished = {Hugging Face Hub},
  url          = {https://huggingface.co/datasets/ezharjan/ViewGraphBench}
}

The references of the baseline methods are listed in the toolkit README.

License

MIT License, Copyright (c) 2026 Aizierjiang Aiersilan. See LICENSE.