Datasets & Metrics
From reproducible BOP comparisons to category generalization, unseen-object onboarding, transparent materials, and hand–object interaction: choose data by task requirements, then interpret results with metrics that match the protocol.
POSE DATASET ATLAS
This catalog follows the task taxonomy of the referenced survey and adds public resources released later or not listed separately. Each record explains what it evaluates, what it requires, and where it is likely to fail; training assets and fixed evaluation benchmarks are listed separately to clarify their intended use.
Quick selection
Showing 40 resources
BOP Classic
LINEMOD-Occluded
Adds ground truth for all modeled objects in one LINEMOD test sequence, creating a classic occlusion stress test.
Recommended for: Single-frame pose under heavy occlusion with known models, and BOP Classic comparisons.
BOP Classic
Industrial parts without distinctive texture or color, with similar shapes / sizes and multiple symmetries.
Recommended for: Symmetry-aware evaluation, missing texture, industrial grasping, and BOP AR.
BOP Classic
MVTec ITODD
Industrial objects captured with high-quality Gray-D sensors. Training / validation ground truth is public, while test ground truth is controlled.
Recommended for: Industrial parts, depth perception, and BOP's controlled test protocol.
BOP Classic
HomebrewedDB
RGB-D data mixing toys, household objects, and industrial parts, captured with Primesense and Kinect sensors.
Recommended for: Validation across known-model RGB-D, different sensors, and mixed object domains.
BOP Classic
YCB-V
A tabletop multi-object video benchmark widely used for known-object RGB / RGB-D 6D pose. The BOP version supplies unified formats and PBR training images.
Recommended for: Classic multi-object RGB-D comparisons; avoid original synthetic training images that leak test poses.
BOP Classic
Two IC-MI objects appear in multiple bin locations under severe occlusion.
Recommended for: Small-scale, heavily occluded bin-picking checks for qualitative performance and protocol validation.
BOP Classic
TUD Light
Training and test sequences of three moving objects under eight lighting conditions.
Recommended for: Controlled lighting-robustness experiments. Its small scale is insufficient as the sole basis for generalization claims.
BOP Classic
LM
A classic single-object 6D pose dataset. Each sequence mainly annotates one target, with relatively mild occlusion.
Recommended for: Basic pipeline reproduction and implementation checks. Scores should not be extrapolated to complex occlusion.
BOP-H3
Purchasable toy grocery objects captured in homes / offices. HOPEv2 extends real multiview evaluation for BOP-H3.
Recommended for: Home robotics, reproducible object procurement, and seen / unseen object detection in BOP-H3.
BOP Classic
Rutgers APC
RGB-D poses of Amazon Picking Challenge products on cluttered warehouse shelves.
Recommended for: Warehouse pick-and-place, textured products, and occlusion.
BOP Classic
Two weakly textured and four textured household objects, with multiple instances and mild occlusion.
Recommended for: Reproducing early RGB-D 6D pose methods and contrasting occlusion difficulty with IC-BIN.
BOP Classic
Toyota Light
Controlled tabletop captures of 21 objects, emphasizing changes in pose, tablecloths, and lighting.
Recommended for: RGB-only comparisons under lighting and background changes.
BOP-H3
Headset-based hand–object tracking data with ground-truth objects, hands, cameras, gaze, and more. BOP uses a quality-checked Clips subset.
Recommended for: Dynamic object onboarding, egocentric hand–object interaction, and BOP-H3 unseen-object tasks.
BOP-H3
A handheld-object dataset. BOP selects a subset with available CAD and adds a further test set.
Recommended for: Hand occlusion, known / unknown model detection, and BOP-H3 tasks.
BOP Industrial
Industrial Parts Dataset
Industrial multiview data captured using HDR and structured light, included in BOP Industrial.
Recommended for: Multiview industrial inspection, cross-view fusion, and the BOP Industrial protocol.
BOP Industrial
XYZ Industrial Bin-picking Dataset
Industrial bin-picking data with accurate ground-truth object poses in the BOP Industrial format.
Recommended for: Multiview bin picking and stress tests for real industrial deployment.
BOP Industrial
A multiview extension of ITODD for single-view and multiview detection of known or unseen object models in BOP Industrial.
Recommended for: Multiview extension experiments for systems already evaluated on ITODD.
Independent benchmark
YCB video sequences for object manipulation and action tasks, retaining occlusions caused by hand–object interaction.
Recommended for: Known-object tracking in interactive manipulation sequences.
Independent benchmark
A large-scale RGB-D dataset of transparent objects, designed to cover failure modes of depth and color observations on transparent materials.
Recommended for: Transparent-object 6D pose estimation, transparent-object depth completion, and robustness across materials.
Independent benchmark
A mixture of real and synthetic data for metal-part pose estimation, emphasizing highly reflective surfaces.
Recommended for: Metal industrial parts, transfer between real and synthetic data, and robustness to reflections.
Independent benchmark
A large-scale stereo RGB dataset for 6D pose estimation, capturing transparent, reflective, and symmetric objects in different environments.
Recommended for: Stereo geometry, transparent and reflective objects, and large-scale instance-level training.
Independent benchmark
NOCS CAMERA
The synthetic component of the NOCS benchmark, using ShapeNet models to construct category-level RGB-D training and evaluation settings.
Recommended for: Category-level 6D pose and scale estimation, and synthetic pretraining; report results together with REAL275.
Independent benchmark
NOCS REAL
The real-data component of the NOCS benchmark, primarily evaluating category-level pose and size on unseen instances.
Recommended for: The main real-world category-generalization benchmark paired with CAMERA25.
Independent benchmark
Represents category-level objects using keypoints and manipulation constraints, focusing on robotic manipulation rather than CAD alignment alone.
Recommended for: Category-level grasping, affordances, and keypoint-driven pose representations.
Independent benchmark
Transparent Object Dataset
A transparent-object dataset with stereo and RGB-D observations, covering category-level pose and size for transparent materials.
Recommended for: Transparent-object category-level methods, stereo input, and scale evaluation.
Independent benchmark
Object-centric data captured with mobile devices, providing object 3D boxes, keypoints, and camera poses for real mobile scenarios.
Recommended for: RGB-only methods, mobile AR, and category-level 3D boxes/9DoF; its annotations do not specify fine-grained poses relative to CAD surfaces.
Independent benchmark
Category-level 6D pose data from real videos, emphasizing natural environments, fast motion, and occlusion.
Recommended for: In-the-wild RGB video, category-level temporal pose estimation, and domain generalization.
Independent benchmark
Category-level data combining photometric and polarization information, including optically challenging transparent and reflective objects.
Recommended for: Polarization/RGB-D methods, category-level pose across materials, and optically challenging cases.
Independent benchmark
Multimodal category-level data from real household scenes, covering everyday, transparent, and reflective objects, with grasp annotations.
Recommended for: Category-level RGB-D/polarization methods, household environments, and grasp-aware perception.
Independent benchmark
A large-vocabulary category-level RGB-D dataset that expands the category count from the traditional NOCS setting to 166 and adds complex backgrounds and occlusion.
Recommended for: Large-vocabulary category-level pose estimation, broader generalization, and symmetry-aware evaluation.
Independent benchmark
BMVC
A small, early articulated-object pose collection included in the survey, mainly useful as a historical comparison with newer articulated-object datasets.
Recommended for: Reproduction of early articulated-pose methods; insufficient on its own to support generalization claims.
Independent benchmark
RBO Dataset of Articulated Objects
Real interactive articulated-object data for studying moving parts, occlusion, and temporal continuity.
Recommended for: Articulation-state estimation, interactive scenes, and visual robotics research.
Independent benchmark
Large-scale egocentric RGB-D interaction data with per-frame annotations of category-level object poses, hand poses, actions, and scenes.
Recommended for: Category-level object pose tracking, hand–object interaction, and embodied temporal tasks.
Independent benchmark
ReArtMix provides large-scale synthetic interaction and articulation data, while ReArtVal provides real validation videos; together they support synthetic training and real-world evaluation.
Recommended for: Generalization across articulated categories, with synthetic training and real-world validation.
Independent benchmark
Articulated-object and hand-contact data collected in a physics simulator through visual teleoperation, with ground-truth part poses and contacts.
Recommended for: Contact priors and joint pose estimation of hands and articulated objects; account for the simulation-to-reality gap.
Independent benchmark
Unseen-object pose data for test-time multiview object modeling and matching across environments.
Recommended for: Model-free methods, test-time modeling, and generalization across environments.
Independent benchmark
A generalizable object pose benchmark that uses reference images instead of a predefined CAD model.
Recommended for: Reference-based unseen-object pose estimation and RGB-only generalization.
Independent benchmark
Builds an object representation from an ordinary phone video before testing, then estimates the pose of the unseen object in cluttered scenes.
Recommended for: Deployable object onboarding, reference-video reconstruction, and model-free pose estimation.
Independent benchmark
A challenging low-texture subset of OnePose for testing unseen-object pose methods that rely only on local features.
Recommended for: Low-texture object onboarding and analysis of generalization failures.
Independent benchmark
MegaPose
Large-scale synthetic training resources for BOP unseen-object tasks, rather than a standalone benchmark with a fixed real-world test set.
Recommended for: FoundationPose/MegaPose-style pretraining and BOP unseen-object tasks; evaluation still requires a specific benchmark.
No matches. Clear the filters or try transparent, occlusion, RGB-D, or unseen objects.
Inclusion criteria: publicly available or explicitly requestable images/videos with pose annotations, or annotations directly supporting pose research. Small historical collections are retained, with their limitations noted in the recommended uses.
General asset libraries containing only 3D models and no pose-task annotations are not listed individually. The sole exception is MegaPose training data, explicitly intended for unseen-object pose pretraining and labeled as training assets. Data licenses, download links, and versions can change; consult the project page before starting experiments.
Metrics
Error functions, threshold rules, and final scores describe different stages of evaluation. First identify where error is measured, then check symmetry handling, occlusion handling, and aggregation.
Each card links to a detail page: start with the quick summary, then explore formulas, diagrams, limitations, BOP protocols, and related papers in the library.
View all metrics →Geodesic Rotation Error
Measures the smallest rotation angle between predicted and ground-truth orientations, usually reported alongside translation error or thresholds such as 5°.
3D Translation Error
Measures the 3D distance between predicted and ground-truth object centers; always check the coordinate frame and whether units are m, mm, or cm.
Rotation–Translation Threshold Accuracy
A prediction succeeds only when both rotation and translation errors fall below their thresholds; 5°2cm, 5°5cm, and 10°5cm are examples.
Average Distance of Model Points
Transforms CAD model points with the predicted and ground-truth poses, then averages their 3D distances. It suits asymmetric objects but penalizes visually equivalent symmetric poses.
Average Closest Point Distance
Uses the nearest-point distance for each model point to reduce ambiguity between equivalent poses of symmetric objects. Commonly used for objects with axial symmetry or repeated structures.
ADD / ADD-S reporting protocol
Selects ADD or ADD-S according to object symmetry, then aggregates results. It does not define a new distance formula; check the object grouping and threshold as well.
Area under the ADD recall curve
Averages recall over a range of error thresholds, providing a more continuous measure than ADD-0.1d alone; the integration limit must match.
2D Projection Error
Compares predicted and ground-truth CAD projections in the image plane, often reported as Proj@5px / Prj-5; particularly useful for RGB-only visual alignment.
Visible Surface Discrepancy
Compares visible surfaces under predicted and ground-truth poses, focusing on what is actually visible in the image to handle ambiguity from occlusion and symmetry.
Maximum Symmetry-Aware Surface Distance
The maximum surface distance after accounting for object symmetries, emphasizing the worst-case deviation in 3D space.
Maximum Symmetry-Aware Projection Distance
Compares projection error in the image plane to show how far a pose error displaces points in the 2D image.
BOP Average Recall
BOP computes recall at multiple thresholds for VSD, MSSD, and MSPD, then aggregates it in stages into AR; AR is not a distance measure.
Average Precision for 6D Detection
When target identities are unknown, BOP detection jointly evaluates detection, confidence ranking, and pose, using AP rather than localization AR.
3D Bounding-box Intersection over Union
Commonly used in category-level tasks that predict size, this metric compares 3D box overlap and depends on position, rotation, and dimensions.
Bidirectional nearest-point shape distance
Uses bidirectional nearest-point distances to evaluate shape reconstruction or completion. It complements pose metrics; the two cannot replace each other.
Reading Results
A method can perform very differently across datasets and metrics. Check the scene conditions before comparing numbers.
For objects with rotational symmetry or repeated structures, standard ADD may penalize reasonable equivalent poses. Also consider ADD-S or BOP symmetry-aware metrics.
RGB-D methods typically achieve better translation and surface alignment, but have different deployment requirements and should be distinguished from RGB-only methods.
BOP AR, traditional ADD(-S) accuracy, projection error, and paper-specific thresholds measure different quantities. Check the protocol before comparing results.
