OASIS 2 Dataset Overview

• OASIS 2 Dataset is a dual-use term: in computer vision, it refers to a large-scale, richly annotated single-image 3D dataset, while in ontology engineering it designates an OWL 2 ontology for agent behaviours.
• The vision dataset provides detailed per-image annotations for reconstructing depth, normals, and occlusion/fold boundaries, leveraging an interactive, multi-stage quality control workflow.
• The ontology distribution models agents’ capabilities and processes using formal OWL/RDF representations, supporting reasoning tasks in role-based behaviour and procedural execution.

Searching arXiv for the cited OASIS papers and any related versioning/context. {"query":"OASIS single image 3D in the wild (Chen et al., 2020) OASIS 2 ontology (Bella et al., 2023)","max_results":10} I’ll look up the exact arXiv records to ground the article in the current literature. The expression “OASIS 2 Dataset” is domain-dependent rather than singular. In computer vision, the relevant resource is OASIS — Open Annotations of Single-Image Surfaces, introduced as a large-scale dataset for single-image 3D in the wild; the associated paper and project site do not define a distinct dataset or version named “OASIS 2,” and the public release is simply called OASIS (Chen et al., 2020). In ontology engineering, by contrast, OASIS 2 denotes Ontology for Agents, Systems and Integration of Services, version 2, a foundational OWL 2 ontology whose downloadable OWL/RDF files, and related example ABox distributions, are sometimes treated as a “dataset” (Bella et al., 2023). The term therefore names either a non-existent second version of the CVPR 2020 vision dataset, or an actual ontology distribution in the semantic-web literature.

1. Terminological scope and versioning

The two uses of the name differ in both object type and scholarly context. In the vision literature, OASIS is a dataset of annotated images for monocular 3D reconstruction. In the ontology literature, OASIS 2 is an OWL 2 ontology for agent behaviours, plans, processes, and procedures.

Term	Referent	Versioning status
OASIS	Open Annotations of Single-Image Surfaces	Publicly released dataset; minor post–camera-ready quality fixes affecting $<10\%$ of images
“OASIS 2” in vision	Purported second version of OASIS	Not defined by the authors
OASIS 2	Ontology for Agents, Systems and Integration of Services, version 2	Actual ontology distribution in OWL/RDF form

For the CVPR 2020 resource, the first public release is associated with “OASIS: A Large-Scale Dataset for Single Image 3D in the Wild” (Chen et al., 2020). The paper states that the publicly released dataset is an improved version relative to the camera-ready, with small post–camera-ready quality fixes affecting fewer than $10\%$ of images, but it does not introduce any “OASIS 2” nomenclature. A common misunderstanding is therefore to read “OASIS 2 Dataset” as a formal second release of the vision benchmark; the source material does not support that interpretation.

In the ontology paper, “A behaviouristic approach to representing processes and procedures in the OASIS 2 ontology”, the phrase OASIS 2 refers to the ontology itself and its OWL/RDF distribution, with related ABox/example distributions available separately (Bella et al., 2023). In that setting, “dataset” typically means the ontology files that can be downloaded and loaded into a reasoner or triplestore.

2. OASIS as a single-image 3D dataset

OASIS is defined as a dataset for single-image 3D in the wild consisting of annotations of detailed 3D geometry for 140,000 images (Chen et al., 2020). The images are random Creative Commons Flickr images with known focal length (from EXIF). The domain is explicitly heterogeneous: indoor, outdoor, natural, and man-made scenes are all represented, and focal lengths vary widely, with most between $1\times$ and $10\times$ the image width.

The dataset is organized around per-image annotations designed to enable pixel-wise surface reconstruction. Annotators provide occlusion boundaries with near/far ordering, and distinguish smooth occlusions from sharp occlusions. They also mark fold boundaries, defined as surface-normal discontinuities where surfaces remain physically connected. For smooth surfaces, they indicate whether a surface is planar or curved; they supply one normal per planar surface and multiple normals on curved surfaces where needed. For planar surfaces, they may additionally label relative normals as parallel, orthogonal, or neither, with automatic adjustment of normals to enforce those relations.

The scene geometry represented by these annotations is intentionally nontrivial. Most annotated regions, more than $90\%$, contain occlusion boundaries, and nearly half contain both occlusion and fold boundaries. Most regions, more than $70\%$, include curved surfaces, not only planes. Many regions also comprise multiple disconnected pieces and complex continuity or smoothness. This suggests that OASIS was constructed to stress the failure modes of planar, indoor-only, or sensor-constrained datasets rather than to regularize them away.

The distributed outputs include RGB images, per-pixel dense depth maps, per-pixel dense normal maps, binary maps for fold and occlusion boundaries with occlusion near/far side labels, planar instance masks and planar parameters, and per-image focal length. The paper does not enumerate specific file formats such as PNG, EXR, or NPZ, does not specify exact dataset size on disk, and does not explicitly specify image resolution; those implementation details are deferred to the project site.

3. Annotation workflow and geometric reconstruction

The annotation pipeline is built around an interactive workflow rather than direct dense labeling. Annotators first select a polygonal region that is required to cover two randomly pre-selected locations. They then mark occlusion and fold lines, annotate normals with a draggable 3D arrow, decide planarity for smooth surfaces, and specify parallel or orthogonal relations for planar pairs. An interactive 3D preview window shows the reconstructed surface mesh, and annotators iterate until satisfied (Chen et al., 2020).

Quality control is multi-stage. Workers complete training on occlusions, folds, normals, and UI usage, and must pass a qualification quiz. Each collected mesh is then subjected to a separate quality verification task in which workers judge whether the reconstructed mesh reflects the image well; only meshes deemed high-quality are accepted. The workforce is mixed: 11% annotations via Amazon Mechanical Turk; the remainder via two annotation companies. The reported average time per image is approximately 305 seconds, indicating a deliberately high annotation effort.

Dense geometry is not hand-labeled directly but derived from the sparse structural annotations. Dense normals are obtained by propagating annotated normals with a smoothness objective, stopping at fold and occlusion boundaries, and solving via least squares (LU factorization) with boundary-aware neighborhood constraints, followed by per-pixel normalization. Dense depth is then recovered by integrating normals, with known focal length, per continuous surface to obtain depth up to per-surface scale. Inter-surface ordering is enforced by a constrained optimization that finds a scale $X_S$ for each surface $S$ so that near/far ordering along sampled occlusion pairs is satisfied; a small epsilon separation and minimum scale eta are used.

This reconstruction strategy imposes a particular scale semantics. Depth is metric only up to a per-surface scale; ordering across surfaces is enforced by scaling, but absolute global scale across the whole image is not guaranteed. Normal coordinates are effectively camera-centric, and the integration pipeline treats them in the camera frame. The authors also clip or snap the z-component of normals to be no smaller than 0.3 during normal integration to avoid excessive depth stretch. A second common misconception is therefore to treat OASIS depth as globally scaled metric ground truth; the paper explicitly cautions against that interpretation.

4. Tasks, metrics, and empirical profile

OASIS provides official splits of train 110K, val 10K, test 20K, applied consistently across tasks (Chen et al., 2020). The supported tasks are single-image depth estimation, surface normal estimation, joint boundary detection and classification, and instance segmentation of planes with plane parameter regression.

The depth protocol is built around Locally Scale-Invariant RMSE (LSIV_RMSE) on 3D point clouds after backprojection with predicted focal length and per-surface scaling and z-translation alignment. WKDR is reported for depth ordering. The evaluation guidance is explicit: models should predict focal length, or use a validation-selected focal length if focal length is not produced; LSIV_RMSE alone is insensitive to cross-surface ordering, so ordering should be evaluated separately via WKDR. The paper also lists common depth metrics such as AbsRel, SqRel, RMSE, RMSE$_{\log}$, thresholded $\delta$ accuracies, and scale-invariant depth error, while emphasizing LSIV_RMSE and WKDR for OASIS.

For depth estimation, representative results are: ResNetD trained on ImageNet + OASIS: LSIV_RMSE $10\%$0; WKDR $10\%$1. The same architecture trained only on OASIS yields LSIV_RMSE $10\%$2; WKDR $10\%$3, and an Hourglass model trained on OASIS yields LSIV_RMSE $10\%$4; WKDR $10\%$5. Off-the-shelf FCRN and Hourglass models trained on other datasets are reported at approximately 0.67 LSIV_RMSE, with WKDR around 38–40%, and are not better than a plane baseline. Approximate human performance is LSIV_RMSE $10\%$6; WKDR $10\%$7.

For surface normals, the reported metrics are mean and median angular error and the percentage of pixels within 11.25°, 22.5°, and 30°. An Hourglass model trained on OASIS achieves mean angular error $10\%$8, median $10\%$9, with 31.23% / 59.45% / 71.77% within the three thresholds, and relative-normal AUC$1\times$0 $1\times$1 and AUC$1\times$2 $1\times$3. Approximate human performance is mean $1\times$4, median $1\times$5, with 44.36% / 76.16% / 85.24% within thresholds, and AUC$1\times$6 $1\times$7 and AUC$1\times$8 $1\times$9. The paper also reports that an Hourglass model trained on OASIS generalizes better to DIODE and ETH3D than models trained on NYU or SNOW.

For fold/occlusion boundary detection, the metrics are ODS, OIS, and AP, with a detection required both to localize the boundary and to classify it correctly as occlusion or fold. An Hourglass model trained on OASIS yields ODS $10\times$0, OIS $10\times$1, AP $10\times$2; HED trained on OASIS yields ODS $10\times$3, OIS $10\times$4, AP $10\times$5; approximate human performance is ODS $10\times$6. For planar instance segmentation, PlanarReconstruction trained on ScanNet only gives AP $10\times$7, AP50 $10\times$8, AP75 $10\times$9; trained on OASIS, AP $90\%$0, AP50 $90\%$1, AP75 $90\%$2; trained on ScanNet + OASIS, AP $90\%$3, AP50 $90\%$4, AP75 $90\%$5; approximate human performance is AP $90\%$6, AP50 $90\%$7, AP75 $90\%$8.

These results support several bounded conclusions already stated by the authors. Training directly on OASIS and predicting focal length improves metric reconstruction in the wild; ImageNet pretraining helps; relative normal relations remain difficult; joint fold-versus-occlusion boundary detection is challenging; and sizable human–machine gaps remain across tasks. The paper also notes that standard normal metrics may underweight foreground geometry, which suggests that current evaluation may not fully capture perceptual quality in complex scenes.

5. OASIS 2 as an OWL 2 ontology distribution

In the ontology literature, OASIS 2 is a foundational OWL 2 ontology for modeling agents, their mental states and capabilities, and their interactions, including plans and action executions (Bella et al., 2023). Its behavioristic orientation is explicit: it emphasizes what agents can do, how those abilities are described, and how they are executed. In this context, the “dataset” is the downloadable ontology distribution, including the ontology files and, in related resources, example ABox individuals.

The pre-extension core model includes classes such as Agent, Behaviour, GoalDescription, TaskDescription, TaskOperator, Action, TaskOperatorArgument, TaskObject, TaskInputParameter, TaskOutputParameter, TemplateThing, BehaviourThing, PlanningThing, ExecutionThing, and PlanExecution. The intended semantics are specific. A Behaviour is a mental state describing agent capabilities and collects goals; a GoalDescription is a container of tasks an agent can achieve; a TaskDescription models atomic operations; TaskOperator refers to the action to be performed; and the various task-object and task-parameter classes specify recipients, inputs, and outputs. Template, behaviour, planning, and execution markers distinguish abstract templates, concrete behaviours of real agents, planned actions, and already performed actions.

Selected properties structure this capability-and-execution model. hasBehaviour links an agent to a behaviour; consistsOfGoalDescription and consistsOfTaskDescription compose behaviours into goals and tasks; dependsOn organizes inter-goal and inter-task dependency; refersExactlyTo, refersAsNewTo, and refersAsInstanceOf connect descriptions to concrete or template-level entities; overloadsBehaviour and related overload properties connect concrete behaviours to templates; drawnBy subproperties trace executions back to the elements they were drawn from; submittedTo subproperties link planned descriptions to the behaviours responsible for realization; and hasExecution subproperties connect plans to their executions.

The 2023 extension adds explicit representations of processes and procedures. Motivated by ISO 9001-2015, the paper distinguishes a process as a set of correlated or interactive activities that transform inputs into outputs, and a procedure as a description of how to carry out all or part of a process. New classes include Process, Procedure, and ProcedureState with two categories: TerminatingProcedureState, containing InitialProcedureState and FinalProcedureState, and NonTerminatingProcedureState. The paper states the representative axiom Procedure $90\%$9 Activity, and introduces ordering and composition properties such as procedureConsistsOfProcedureState, procedureConsistsOfInitialProcedureState, procedureConsistsOfFinalProcedureState, hasNext, hasNextNonTerminatingProcedureState, hasFinalProcedureState, and hasNextProcedure.

Each procedure state is linked to the planned behaviour that describes how to perform it by isDescribedBy, and may be connected to an Event by triggersEvent. The ontology also models roles and behaviour provisioning: Role can provide behaviours via providesBehaviour; agents acquire those behaviours by playRole; RoleType provides domain-specific classification; and DeprecatedThing marks deprecated roles. The result is an ontology that integrates capabilities, plans, ordered procedure steps, executions, roles, and events within a single agent-centric formalization.

6. Access, usage patterns, and interpretive cautions

The two resources are accessed and used differently. The vision dataset is available from the OASIS project site, and its use is conditioned by the fact that the images derive from Flickr CC; licensing and redistribution therefore follow source constraints, and the latest terms should be checked on the project page (Chen et al., 2020). The ontology distribution is available from the OASIS 2 v2 ontologies repository, with separate links for the OASIS-Abox ontology and Ether-OASIS; the paper is CC BY 4.0, but it does not explicitly state the ontology’s license, so the repository should be consulted for licensing details (Bella et al., 2023).

The ontology paper does not advertise a public SPARQL endpoint. Typical usage is to load the OWL files into a reasoner or triplestore, for example Protégé with HermiT, Pellet, or FaCT++. The intended reasoning tasks include checking consistency of processes and procedures, role-based behaviour provisioning, and tracing plan-to-execution links through the drawnBy family of properties. The paper’s case study is a resource lock process with three procedure states—request lock, modify resource, and release lock—each linked by isDescribedBy to a planned behaviour, and realized through corresponding plan executions connected by hasPlanExecution and planExecutionDrawnBy.

For the vision dataset, the principal methodological caution is to align predictions per-surface and report ordering separately. The paper explicitly states that OASIS depth should not be treated as globally scaled metric ground truth, that LSIV_RMSE is invariant to per-surface scale and translation in $70\%$0, and that WKDR reporting is important because LSIV_RMSE does not enforce ordering between disconnected surfaces. For the ontology, the principal interpretive caution is terminological: “dataset” refers to the OWL files and possibly associated ABox examples, not to a benchmark with empirical performance metrics. The 2023 ontology paper itself presents no empirical performance metrics; its evaluation is by formalization, competency questions, example RDF/Turtle fragments, and standard OWL 2 reasoning practice.

Taken together, the two literatures show that “OASIS 2 Dataset” is not a single canonical object. In single-image 3D, the authoritative dataset is OASIS, not “OASIS 2.” In knowledge representation, OASIS 2 is an actual ontology distribution for agent behaviours, processes, procedures, and executions. Any precise use of the term therefore depends on whether the intended referent is a computer-vision dataset or an OWL 2 ontology package.