ncsu-digital-manufacturing/teeth_retainer_v2

Teeth Retainer v2 — Domain Randomization

Dataset summary

Teeth Retainer v2 is an image dataset with bounding-box and 2D keypoint annotations for dental-retainer perception experiments. It organizes synthetic domain-randomization variants, physical-image data, a mixed-data condition, and synthetic-training/real-validation conditions into separate folders. It supports studying how changes in simulated appearance and camera settings affect detection and keypoint localization, including transfer to physical images.

The repository is ncsu-digital-manufacturing/teeth_retainer_v2. This is the dataset repository; a separately hosted simulation asset with the same short name is not part of this dataset inventory.

The inspected snapshot contains 17 dataset variants, 2,368 image file entries, and 2,366 matching-format annotation files. These are totals across variant folders, not a count of unique scenes or independent observations. Related variants may reuse images or labels. The repository also contains 12 class-distribution plot images, which are excluded from the image totals above.

Intended uses

• Train and evaluate bounding-box detectors and 2D keypoint estimators for retainers and keyhole features.
• Compare individual domain-randomization factors and selected combinations.
• Compare physical-only, synthetic-only, mixed-data, and synthetic-to-real experimental conditions.
• Research and teaching in robotic perception and digital manufacturing.

No benchmark scores or demonstrated real-world transfer performance are supplied in the inspected files. The dataset does not establish clinical suitability or reliable physical robot operation.

Dataset organization and splits

Every variant has a data.yaml and split folders containing images/ and labels/. The on-disk validation folder is named valid; YAML configurations refer to it using the val key.

Variant folder	Randomization or experimental condition indicated by its name	Train images	Valid images	Test images
phy_data	Physical-image baseline	105	29*	—
data_mixed_obj_focal_background_light_random_real_140_imgs	Mixed physical/synthetic data with object, focal, background, and light variation	105	28	—
sim_data_new_background_random_140_imgs	Background	105	28	7
sim_data_new_camera_angle_random_140_imgs	Camera angle	105	28	7
sim_data_new_camera_params_random_140_imgs	Camera parameters	105	28	7
sim_data_new_distance_random_140_imgs	Distance	105	28	7
sim_data_new_exposure_random_140_imgs	Exposure	105	28	7
sim_data_new_focal_random_140_imgs	Focal setting	105	28	7
sim_data_new_focal_random_bigger_range_0.5_2.5_140_imgs	Wider focal setting; name includes 0.5_2.5	105	28	7
sim_data_new_light_random_140_imgs	Lighting	105	28	7
sim_data_new_principal_point_random_140_imgs	Camera principal point	105	28	7
sim_data_obj_randomization_140_imgs	Object randomization	105	28	7
sim_data_obj_randomization_background_random_140_imgs	Object + background	105	28	7
sim_data_obj_randomization_focal_random_140_imgs	Object + focal setting	105	28	7
sim_data_obj_randomization_light_random_140_imgs	Object + lighting	105	28	7
sim_train_real_val_focal_random	Synthetic training, physical validation; focal condition	105	29*	7
sim_train_real_val_obj_randomization	Synthetic training, physical validation; object condition	105	28	7

— means no split folder was observed. * marks a validation folder with 29 image files but only 28 label files. Each contains an unmatched image named frame0343_png.rf.ee6734f2358f3a9940e3a07997c9e405 (Copy).jpg.

The 13 synthetic randomization variants each contain 140 images with a 105/28/7 split (75%/20%/5%). Despite its _140_imgs suffix, the mixed-data folder contains 133 images in the inspected snapshot. All split counts include images whose corresponding annotation file is empty.

Example layout:

teeth_retainer_v2/
├── class_count/                         # 12 distribution plots
├── phy_data/
│   ├── data.yaml
│   ├── train/{images,labels}/
│   └── valid/{images,labels}/
├── sim_data_new_focal_random_140_imgs/
│   ├── data.yaml
│   ├── train/{images,labels}/
│   ├── valid/{images,labels}/
│   └── test/{images,labels}/
└── ...                                 # Other variants listed above

Training and validation folders also contain labels.cache files. These are derived cache artifacts, not ground-truth annotations; regenerate them when paths or labels change.

Images

Synthetic variant images use PNG; physical images use JPEG. The mixed condition has 83 PNG and 22 JPEG files in training, and 22 PNG and 6 JPEG files in validation. The two sim_train_real_val_* conditions have PNG training/test images and JPEG validation images.

Representative image headers from every split showed 1280 × 720 pixels for synthetic images and 2048 × 1152 pixels for physical images. These dimensions were sampled, not exhaustively checked. The supplied annotation script independently sets its synthetic render resolution to 1280 × 720. Preserve each image's actual dimensions when converting normalized annotations to pixels. Image extensions alone do not establish capture provenance.

Both sampled resolutions have a 16:9 aspect ratio. Different source resolutions are compatible with normalized pose labels when images and annotations are transformed consistently. For cross-domain evaluation, use the same model input size and preprocessing; compare keypoint errors in normalized coordinates or at a common image resolution rather than comparing raw source-pixel errors.

Classes and annotations

All 17 data.yaml files specify the same class order and keypoint structure:

nc: 3
names: ['keyhole_bottom', 'keyhole_top', 'retainer']
kpt_shape: [13, 3]
flip_idx: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]

Class ID	Class name	Annotated instances across all variants
0	keyhole_bottom	7,175
1	keyhole_top	5,762
2	retainer	5,790

There are 18,727 annotated object instances across the variant folders. Reused data may contribute repeatedly to these totals.

An image and its label file share a filename stem. Labels use a YOLO-style pose format, with one object instance per nonempty line and 44 whitespace-separated numeric values:

class_id cx cy width height x0 y0 v0 x1 y1 v1 ... x12 y12 v12

• cx, cy, width, and height describe a normalized center-based bounding box.
• Each keypoint has normalized image coordinates (x, y) and a visibility value v.
• Under the standard pose-label convention, 0 denotes an unannotated point, 1 an annotated but non-visible point, and 2 a visible point. The inspected labels contain values 0 and 2; they do not independently verify physical visibility.
• The supplied script defines retainer landmarks keypoint_0 through keypoint_12, top-keyhole landmarks keypoint_13 through keypoint_18, and bottom-keyhole landmarks keypoint_19 through keypoint_24. Each keyhole group occupies slots 0–5 of its own YOLO row; slots 6–12 are (0, 0, 0) padding. Retainer rows use slots 0–12. These export rules agree with sampled dataset rows, but the historical generating revision is unverified.
• Landmark coordinates are hard-coded in the script, scaled by 0.001, and mapped from a CloudCompare frame into the USD mesh frame using a fixed rotation and translation. The code then transforms them through the rigid body's world pose and projects them using the render annotator's view/projection matrices and resolution. The source identifies a scan-specific registration; its accuracy was not revalidated for this card. Semantic names for individual landmark positions and a landmark diagram are not supplied.
• Actual dictionary/export class IDs are 0=keyhole_bottom, 1=keyhole_top, 2=retainer, matching the dataset YAML. An introductory script comment lists a different mapping; the executable mapping is the basis for this card.
• The supplied flip_idx is the identity permutation. Confirm landmark symmetry and ordering before enabling horizontal-flip augmentation.

Normalized coordinates can be converted using x_pixels = x * image_width and y_pixels = y * image_height. For the bounding box, xmin = (cx - width / 2) * image_width, with analogous formulas for the other edges.

Domain-randomization methodology

Source and historical scope

The creator supplied replicator_autoannotation.py from models/domain_randomization/DM_Auto_Annotation/, inspected on November 10, 2026. The descriptions below document that script's current implementation. It is external to the inspected dataset inventory and is not proven to be the revision that generated every existing variant. Folder names identify intended historical conditions; exact historical settings require run records.

The script runs as an asynchronous task in the NVIDIA Isaac Sim Script Editor, using Replicator, USD, and PhysX APIs with an already-open stage. It expects the conveyor at /World/ConveyorBelt_A06 and ZED X camera prims under /World/stand/ZED_X_01/base_link/ZED_X/. It downloads meshes/retainer_scan/retainer_scan.usdc from the separate Hugging Face asset repository ncsu-digital-manufacturing/teeth_retainer_v2 at mutable revision main. Exact Isaac Sim version, renderer configuration, source-stage revision, and asset commit are not pinned.

Implemented randomization

Ranges below apply to the supplied script, not automatically to all folders in this release. Uniform sampling is used except where stated otherwise.

Factor	Current implementation
Object count	Integer count from 0 through 16 per scene, inclusive. This varies independently of the selected camera/appearance mode.
Object placement and orientation	Spawn X within ±0.15 and Y within ±0.05 of the conveyor bounding-box center, in stage units; Z is conveyor top + 0.10 + index * 0.10. Initial scene orientations use a uniform-quaternion sampling construction, then evolve under physics. These offsets assume the scene's unit convention; they are not converted using stage meters-per-unit.
Focal length	Shared multiplier from 0.50 to 2.50 applied to each ZED camera's baseline focal length. This is a dimensionless scale, not an absolute focal length or focus-distance range. It clarifies the current implementation associated with the 0.5_2.5 descriptor, without proving historical folder settings.
Camera angle	Additional rig X and Y Euler rotations each from −5° to +5°; added Z rotation is 0°.
Exposure	Post-render RGB adjustment from −1 to +1 EV: decode sRGB to linear RGB, multiply by 2**EV, clip, and re-encode sRGB. This does not simulate sensor exposure time, motion blur, or sensor noise.
Principal point	Shared horizontal and vertical shifts, each from −0.10 to +0.10 times the corresponding camera aperture, added to baseline aperture offsets.
Camera distance	Rig translation along world Z, nominally from −0.50 to +0.50 m. The lower limit is raised as needed so the lower camera remains at least 0.10 m above the conveyor top. This is a height-offset distribution, not a direct camera-to-object distance distribution.
Lighting	One authored distant light. Dim/normal/bright categories have probabilities 25%/50%/25%; intensity is sampled within 50–400, 700–1300, or 1800–3500 respectively in authored USD light-intensity values. Color temperature is 4500–6500 K; rotation X is −65° to −25° and Z is −35° to +35°. These intensity values are not calibrated illuminance measurements.
Background	Conveyor-belt diffuse color selected uniformly from 22 RGB triplets. The belt uses UsdPreviewSurface with roughness 0.55. This implementation varies belt color rather than replacing entire backgrounds or sampling image textures.

RUN_EFFECTS includes baseline, individual focal/angle/exposure/distance/principal-point/lighting/background modes, camera_only, and combined. camera_only enables focal, angle, exposure, distance, and principal-point effects; combined adds lighting and background. The script also lists along_belt and across_belt, but no corresponding effect logic was found, and the latter's effect token has a trailing comma inside the string. These entries should not be treated as implemented randomization modes.

Object count, placement, and orientation vary even in baseline; it is a baseline for the optional camera/appearance effects. The script does not establish geometry or retainer-material randomization, or an exact correspondence between historical obj_randomization folders and this version's modes.

Capture, annotation, and selection

The active capture loop advances physics for 1.0 simulation second before pausing and rendering; capture uses eight render subframes with zero time advance. Although settling thresholds and a settling helper are defined, the active loop does not call that helper or require a settled state. The constants for 600 maximum physics frames and 60 settled frames therefore do not establish that captures are settled.

RGB, camera parameters, image-plane depth, and tight 2D boxes are read from the left camera's render product at 1280 × 720. A right render product is created, but this script's export path does not save right-camera images; it does not establish a stereo-pair dataset.

• Retainer boxes: use the Replicator tight 2D box matched to the object's prim path. Keyhole boxes: use the projected, accepted landmark envelope with 3-pixel padding and image-bound clipping. These box definitions differ and can affect detection comparisons with manually drawn physical-image boxes.
• Visibility: a point must be projected in frame, have positive camera depth, belong to the allowed camera-facing group, and match a valid depth value within 3 mm in a radius-one-pixel neighborhood. The script compares converted point depth to annotator depth; this card does not independently validate that test or its tolerance.
• Facing groups: a normalized dot-product threshold of ±0.05 separates top, bottom, and edge views. Top-facing objects require all 13 retainer and all six top-keyhole points to pass; bottom-facing objects require all six bottom-keyhole points. Edge views and objects failing the applicable full-landmark requirement are made invisible, followed by another render/filter pass until the remaining objects pass.
• YOLO output: accepted points are normalized by the configured image width/height and receive visibility 2; unused or rejected points receive (0, 0, 0). The exporter does not emit visibility 1. Empty scenes can deliberately produce empty label files in this implementation, although the origin of each historical empty label remains unverified.

The visibility-based object removal creates a preference for objects with fully usable landmarks and reduces occlusion coverage. It hides rejected objects rather than merely omitting their annotations. Historical variants must not be assumed to use this same filtering policy without their generating revision.

Reproducibility and output

Current top-level settings are RUN_MODE="focal", NUM_SCENES=1, and RANDOM_SEED=None. The split helper requires at least three images, so this one-scene configuration cannot complete the automatic train/valid/test split. It must be changed for a complete generation run; 140 scenes yield the 105/28/7 proportions observed in the synthetic folders.

The split helper sorts image paths, pairs labels, shuffles with a local seed of 42, assigns rounded 20% validation and 5% test counts (at least one each), and uses the remainder for training. It copies files into split folders and retains the raw outputs. This explains the supplied script's splitting method, but does not prove the historical splits were produced by it.

Scene/light/background random generators use RANDOM_SEED or derived seeds. Camera parameters use a string seed containing the seed value, mode, and scene index, even when the seed value is None; a fixed camera sequence does not imply deterministic physics or object placement. Initial asset setup also uses the global Python random generator. Record the full code, environment, stage, asset revision, and run settings rather than relying on a single seed for exact reproduction.

Runs are written to a unique timestamped directory under TEETH_DATASET_DIR, defaulting to ~/datasets/teeth_retainer. The script writes raw images/, YOLO labels/, per-frame json/, split copies, and config/data.yaml. JSON includes camera/object poses, landmark world/camera/pixel coordinates, visibility flags, and box methods. It does not serialize a complete record of sampled randomization settings or the full camera matrices/intrinsics. No per-frame JSON, generator script, or run manifest was present in the inspected dataset inventory.

Physical capture procedure, physical-camera calibration, manual annotation workflow, exact historical generation settings, and annotation review remain unspecified.

Loading and use

Download or clone the dataset repository with all image and label files, preserving the variant folders. If using Git, retrieve any large-file-managed content as well. Select a single variant for an experiment rather than merging the entire repository implicitly.

The supplied YAML files contain ../train/images, ../valid/images, and ../test/images, with no explicit dataset root. Resolve these paths for your loader; they may point outside the selected variant. phy_data and the mixed-data condition also declare a test path despite having no observed test folder.

For example, create a separate configuration with an absolute root, keeping the original files intact:

from pathlib import Path

dataset_root = Path(r"Z:\digital_twins_dev\data\domain_randomization\teeth_retainer_v2")
variant = dataset_root / "sim_data_new_focal_random_140_imgs"
config = Path.cwd() / "teeth_retainer_focal.local.yaml"

lines = [
    f"path: '{variant.resolve().as_posix()}'",
    "train: train/images",
    "val: valid/images",
]
if (variant / "test" / "images").is_dir():
    lines.append("test: test/images")
lines += [
    "nc: 3",
    "names: ['keyhole_bottom', 'keyhole_top', 'retainer']",
    "kpt_shape: [13, 3]",
    "flip_idx: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]",
]
config.write_text("\n".join(lines) + "\n", encoding="utf-8")
print(config)

Use an annotation-aware pose loader to read the images and corresponding label files. No Hugging Face Datasets loading script or tabular image/annotation manifest is included; automatic image loading alone should not be assumed to attach the pose labels.

Quality checks and known limitations

Dataset inventory and label checks were performed on November 9, 2026; the supplied generation script was reviewed and this card updated on November 10, 2026. Split counts and filename pairing were checked against the Git-tracked file inventory. Every one of the 2,366 tracked label files was read and checked for numeric rows, 44-value row length, finite values, class IDs in 0–2, bounding-box components in [0, 1], valid visibility codes, and active keypoint coordinates in [0, 1]. These checks found no violations. They do not establish annotation accuracy, box containment, correct landmark ordering, or correct visibility.

Known limitations include:

• Unmatched validation images: the two extra (Copy).jpg images noted above have no same-stem label. Review or exclude them explicitly rather than silently interpreting them as negative examples.
• Empty labels: 134 images have an existing but empty label file. These may be intentional negative examples or missed annotations; their status has not been confirmed by visual review.
• Small evaluation splits: seven test images per synthetic condition provide limited evidence about generalization.
• Potential overlap: uniqueness, shared source frames, and duplicate content across variants and splits have not been audited. Do not combine folders and resplit blindly; group by source scene or capture sequence when establishing independent evaluation data.
• Transfer evaluation: the sim_train_real_val_* folders provide physical validation images, while sampled test images are synthetic. Their test split should not be reported as an independent real-world test set. Keep physical validation data separate from any final real-world holdout used to claim transfer performance.
• Coverage: representation of different retainers, materials, workcells, camera systems, and capture conditions is undocumented. Parameter names alone do not demonstrate broad domain coverage.
• Validation scope: full image decoding, image-label overlays, label correctness, content deduplication, and model training were not performed. Included distribution plots were not independently reconciled against the labels.

Provenance, privacy, and responsible use

Repository namespace: ncsu-digital-manufacturing. Individual authors, funding, capture dates, annotation personnel, and detailed source provenance are not specified in the inspected repository. The physical/synthetic interpretation follows the folder naming and observed file organization. The supplied script provides implementation details for a synthetic-generation workflow, but its relationship to each historical variant and the physical acquisition records remain unverified.

If retainer geometry or imagery derives from an individual's dental records, establish the relevant consent and redistribution permissions. No privacy or de-identification review is established by this card. Use the data for research with independent validation of any clinical, manufacturing, or robotic application.

License and citation

No license file or explicit dataset license was found in the inspected tracked inventory. This card therefore does not assign a license or imply permission to redistribute source images, scans, or third-party assets. The dataset owner should specify the applicable license and provenance before publication or reuse requiring those rights.

No associated paper, DOI, or preferred citation was supplied. Until an official citation is provided, identify the dataset repository, the exact revision used, and the access date in experimental reports.

Dataset-card format

This file follows the Hugging Face convention of a Markdown README.md with YAML metadata; see the official dataset-card documentation. Missing provenance and licensing details are stated explicitly rather than inferred.

Generation reference: externally supplied models/domain_randomization/DM_Auto_Annotation/replicator_autoannotation.py, reviewed November 10, 2026. SHA-256: 387a31b6ff9f057ac1e2579400df10e4a566a9d729c52b0bd2e261751281e94f. This identifies the reviewed file; it is not a verified generating revision for the dataset.