vanessabajaj/aruco-synthetic-detection

Synthetic ArUco Marker Detection Dataset

A synthetic dataset for training and benchmarking object detectors on the task of localizing ArUco fiducial markers (single class: aruco_marker). Markers are composited onto real-world background photos with randomized geometric and photometric augmentations, and ground-truth bounding boxes are computed automatically.

The dataset was built to support a controlled comparison of classical, CNN-based, and transformer-based detectors under graded image-degradation conditions.

Contents

Split	Images	Boxes	Negatives (0 markers)
train	10,000	9,983	1,260
val	1,500	1,486	199

• Format: COCO detection format.
• Category: single class, id 1 = aruco_marker.
• Negatives: ~12–13% of images contain no marker, to penalise false positives.

Structure

train2017/ # composited training images val2017/ # composited validation images annotations/ instances_train2017.json # COCO annotations (train) instances_val2017.json # COCO annotations (val)

How it was generated

ArUco markers (cv2.aruco, DICT_4X4_50) are rendered and alpha-blended onto background photos, with per-marker randomized rotation, perspective warp, scale, brightness/contrast jitter, blur, shadows, and partial occlusion. The enclosing axis-aligned bounding box is derived from the transformed marker corners. A fraction of images are left marker-free as negatives.

Difficulty metadata (for sliced evaluation)

Beyond the standard COCO fields, each image and annotation carries extra metadata so results can be broken down by condition:

• Each image has difficulty_tags — a subset of ["blur", "low_contrast", "occlusion", "small_scale", "rotation_perspective"] — and an aug_params dict (blur type, contrast, brightness, per-marker parameters).
• Each annotation has its own aug_params (angle, perspective, scale, occlusion), enabling per-box slicing.

These extra fields are ignored by standard COCO loaders (e.g. pycocotools), so the dataset remains a drop-in COCO dataset. The boolean tag thresholds (SMALL_SCALE_FRAC=0.15, PERSPECTIVE_RATIO=0.09, LOW_CONTRAST_THRESH=1.0) are used only for grouping images into analysis buckets and do not affect the augmentation ranges applied during generation.

Intended use

Training and benchmarking object detectors for fiducial-marker localization. Note that this dataset covers detection only (marker presence + bounding box); decoding a marker's ID is a separate downstream step (e.g. classical cv2.aruco decoding).

Citation

If you use this dataset, please cite this repository.