NoeFontana/locus-tag-bench

Locus-Tag Bench

Synthetic benchmark suite for fiducial-tag detection and camera calibration, rendered with render-tag.

Each config corresponds to one campaign (board family × resolution, or a lighting/sensor variant). Images are rendered in Blender Cycles with ground-truth geometry recovered directly from the scene — no detector-in-the-loop, no human labels.

Configs

Config	Purpose	Images	Board/Tag	Resolution
locus_v1_tag36h11_640x480	Detection (low-res)	50	tag36h11	640×480
locus_v1_tag36h11_1280x720	Detection	50	tag36h11	1280×720
locus_v1_tag36h11_1920x1080	Detection	50	tag36h11	1920×1080
locus_v1_tag36h11_3840x2160	Detection (4K)	50	tag36h11	3840×2160
locus_v1_tag16h5_1920x1080	Detection (dense fam)	100	tag16h5	1920×1080
locus_v1_high_iso_1920x1080	Sensor drift (noise)	50	tag36h11	1920×1080
locus_v1_low_key_1920x1080	Lighting (low-key)	50	tag36h11	1920×1080
locus_v1_raw_pipeline_1920x1080	Raw-like pipeline	50	tag36h11	1920×1080
aprilgrid_golden_v1_1920x1080	Calibration (baseline)	150	AprilGrid 6×6	1920×1080
aprilgrid_distortion_brown_conrady_v1_1920x1080	Calibration (Brown-Conrady)	50	AprilGrid 6×6	1920×1080
aprilgrid_distortion_kannala_brandt_v1_1920x1080	Calibration (KB fisheye)	50	AprilGrid 6×6	1920×1080
charuco_golden_v1_1920x1080	Calibration (ChArUco)	150	ChArUco 6×6	1920×1080

Quick start

from datasets import load_dataset

# Load a single config
ds = load_dataset("NoeFontana/locus-tag-bench", "locus_v1_tag36h11_1920x1080", split="train")
print(ds[0]["image"])          # PIL image
print(ds[0]["tag_id"])         # list[int]  — one entry per detection
print(ds[0]["corners"])        # list[list[[x, y]]] — 4 corners per TAG, variable for BOARD

Row layout

Each row = one image, with per-detection fields aggregated into parallel sequences of length N (number of detections in the image).

Field	Type	Notes
image	PIL.Image	RGB render
image_id	str	scene_XXXX_cam_XXXX
scene_id, camera_idx	int32	Parsed from image_id
tag_id	list[int32]	-1 for BOARD records
tag_family	list[str]	e.g. tag36h11, aprilgrid_6x6, charuco_6x6
corners	list[list[[x, y]]]	4 per TAG; variable per BOARD (often 1 sentinel point)
corners_visibility	list[list[int8]]	Ternary flags (see below). Same shape as corners.
keypoints	list[list[[x, y]]]	Inner saddle/intersection points — only populated for BOARD records (e.g. the 5×5 ChArUco inner grid). Empty for TAG rows.
keypoints_visibility	list[list[int8]]	Ternary flags aligned with keypoints.
eval_margin_px	list[int32]	Per-record margin used for visibility tagging (see below)
eval_complete	list[bool] | null	TAG-only rollup; null for BOARD records
distance, angle_of_incidence, pixel_area, occlusion_ratio, ppm	list[float32]	Photometric/geometric per-tag metadata
position	list[[x,y,z]]	World-space tag centroid
rotation_quaternion	list[[w,x,y,z]]	Tag orientation (w-first)
metadata	list[str]	JSON-encoded per-record extras

eval_margin_px — what it means

Many evaluations need to discount tags/corners that sit right at the sensor edge (partial occlusion, distortion fringing, border interpolation artifacts). Each record carries an eval_margin_px scalar; visibility is tagged ternary on a per-point basis:

• 0 — OUT_OF_FRAME (sentinel (-1, -1) — point not projected)
• 1 — MARGIN_TRUNCATED (in image, but inside the margin_px edge band)
• 2 — VISIBLE (strictly inside [margin, W-margin) × [margin, H-margin))

Calibration configs use eval_margin_px = 5. Detection configs (locus_v1_*) use eval_margin_px = 0 (full image is valid).

Corners vs keypoints asymmetry

	corners (tag boundary)	keypoints (inner saddles)
Count	4 per TAG (variable for BOARD)	0 for TAG, variable for BOARD
Visibility field	corners_visibility	keypoints_visibility
Rollup	eval_complete = all(v == 2)	none — filter per-point
eval_complete behavior	Set to bool for TAG records	null for BOARD records (consumer must examine keypoints_visibility)

Practical filter:

import numpy as np

row = ds[0]
for i in range(len(row["tag_id"])):
    cv = np.array(row["corners_visibility"][i])
    kv = np.array(row["keypoints_visibility"][i])
    # Use only fully in-margin corners
    visible_corners = cv == 2                               # (K,) bool
    # Use only fully in-margin saddles (BOARD only)
    visible_saddles = kv == 2 if len(kv) else np.zeros(0, dtype=bool)

Sidecar files (per config)

Uploaded alongside each config's parquet under {config}/:

File	Contents	When to use
rich_truth.json	Authoritative structured ground truth (versioned envelope + evaluation_context with photometric_margin_px, truncation_policy). Carries the ternary corners_visibility / keypoints_visibility arrays and per-record eval_margin_px.	Primary source of truth for evaluation; use when you want the full detection metadata, not just the image-centric parquet view.
coco_labels.json	Standard COCO-format annotations (keypoints flattened as [x, y, v, …] with COCO v ∈ {0, 1, 2}, num_keypoints counts only v=2).	COCO-compatible pipelines; corners + keypoints are concatenated into a single keypoints list per annotation.
ground_truth.csv	Flat per-detection CSV, one row per detection.	Quick inspection, pandas workflows.
provenance.json	Per-image SceneRecipe (rendering spec: camera, lighting, board, seed).	Deterministic re-rendering; debugging.
manifest.json	Campaign-level metadata (job id, preset, git SHA).	Reproducibility audit.
job_spec.json	Exact experiment YAML used to launch the job.	Reproduce the campaign.
checksums.json	SHA-256 per output file.	Integrity verification.

Load any sidecar:

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    "NoeFontana/locus-tag-bench",
    "aprilgrid_golden_v1_1920x1080/rich_truth.json",
    repo_type="dataset",
)
truth = json.loads(open(path).read())
print(truth["evaluation_context"])
# -> {'photometric_margin_px': 5, 'truncation_policy': 'ternary_visibility'}

Record types (rich_truth only)

Each detection in rich_truth.json has record_type ∈ {"TAG", "BOARD"}:

• TAG — one apriltag. corners has 4 points, keypoints is empty, eval_complete is a bool.
• BOARD — aggregate calibration-board record. corners holds the board anchor (often a single center point), keypoints holds the inner feature grid (e.g. 25 ChArUco saddles), eval_complete is absent/null — use per-keypoint flags.

For the 3 AprilGrid calibration configs, each image emits 1 BOARD + N TAG records (N = number of apriltags on the board). For ChArUco, each image emits 1 BOARD (with 25 saddle keypoints) + 36 TAG records.

Generation

Rendered with render-tag — a Blender-Cycles-backed synthetic-data pipeline with strict host/backend separation and schema-validated SceneRecipe JSON. See provenance.json + job_spec.json in each config for the exact rendering spec.

License

MIT — see the generator repo for the full license text. Images are synthetic (no real subjects, no scraped data).

Citation

@software{locus_tag_bench,
  author  = {Noé Fontana},
  title   = {Locus-Tag Bench: Synthetic benchmarks for fiducial detection and camera calibration},
  url     = {https://huggingface.co/datasets/NoeFontana/locus-tag-bench},
  year    = {2026}
}