REAL275: 6D Pose Benchmark for NOCS

• REAL275 is a real-image benchmark for category-level 6D pose and size estimation, providing canonical annotations for six everyday object categories.
• It supports various methodologies including RGB-only, depth-only, and open-vocabulary protocols to tackle clutter, occlusion, and intra-class variability.
• Evaluation protocols employ metrics like IoU thresholds and rotation–translation errors along with symmetry considerations to assess pose estimation accuracy.

Searching arXiv for recent REAL275-related papers to ground the article. arxiv_search(query="REAL275 NOCS category-level 6D pose estimation", max_results=10, sort_by="submittedDate") REAL275 is the real-image benchmark released with NOCS for category-level 6D object pose and size estimation. Across the literature, it is described as the real-world subset, or real-world test subset, of NOCS, and it is used to evaluate real-world scenes with clutter, occlusion, and intra-class variability across six everyday object categories: bottle, bowl, camera, can, laptop, and mug (Bethell et al., 2024, Li et al., 20 Aug 2025, Lin et al., 2024). Because it provides category-level pose, translation, and size annotations in a canonical space while also exposing real RGB-D sensing and challenging scene composition, REAL275 has become a central benchmark not only for single-view category-level pose estimation, but also for depth-only estimation, RGB-only estimation, tracking, open-vocabulary relative pose, model-free novel-object pose, and zero-shot generalization protocols (Weng et al., 2021, Corsetti et al., 2023, Qu et al., 6 Jan 2025).

1. NOCS provenance and dataset composition

REAL275 is consistently situated within the NOCS framework. Several papers describe it as the real-world counterpart of the NOCS benchmark or as the official real test split, with category-level annotations for 6D pose and object size in Normalized Object Coordinate Space (Bethell et al., 2024, Liu et al., 2023). The benchmark covers six rigid categories—bottle, bowl, camera, can, laptop, and mug—and is designed to test generalization to unseen instances within known categories (Liu et al., 2023, Weng et al., 2021).

The dataset is commonly summarized as containing approximately 4.3K training images from 7 scenes and approximately 2.75K testing images from 6 scenes (Bethell et al., 2024, Lin et al., 2024). Other summaries describe the real split as 13 scenes and 7,000 images overall, with 2,750 images from six scenes reserved for validation or testing (Yang et al., 9 Jul 2025). Multiple papers also emphasize that the test set contains three physical or novel object instances per category, making REAL275 a benchmark for category-level rather than instance-level generalization (Zhang et al., 21 Apr 2025, Corsetti et al., 2024).

REAL275 is repeatedly characterized as a cluttered indoor RGB-D benchmark. It has been described as being captured with a Structure Sensor in cluttered indoor scenes, as containing real-world backgrounds and clutter, and as exhibiting viewpoint variation, occlusion, and significant intra-class variation, particularly for cameras and mugs (Yang et al., 9 Jul 2025, Liu et al., 2022, Bethell et al., 2024). This combination of real sensor data and category-level annotation is what allows the benchmark to support both canonical NOCS evaluations and a wide range of later derived protocols.

2. Task settings built on REAL275

The original use of REAL275 is category-level 6D or 9DoF object pose and size estimation from RGB-D observations. In this setting, methods estimate rotation $R \in SO(3)$, translation $t \in \mathbb{R}^3$, and either isotropic or anisotropic size $s$ for previously unseen instances from the six NOCS categories (Liu et al., 2023, Agarwal et al., 29 Mar 2026). Many works build object point clouds by back-projecting masked depth using camera intrinsics, then sample either 1,024 or 1,028 points per instance for downstream processing (Lin et al., 2024, Agarwal et al., 29 Mar 2026).

REAL275 has also become a benchmark for depth-only pipelines. CLIPose uses depth-only input at inference while building point clouds from REAL275 RGB-D using camera intrinsics and instance segmentation (Lin et al., 2024). HRC-Pose likewise uses the depth stream, with instance masks obtained from an RGB-based detector and segmented depth back-projected to point clouds uniformly sampled to $N_p = 1024$ points (Li et al., 20 Aug 2025).

A separate line of work uses REAL275 for RGB-only category-level pose estimation. The diffusion-based method in "Category Level 6D Object Pose Estimation from a Single RGB Image using Diffusion" eliminates both CAD models and metric depth, using only a single RGB image together with relative depth and surface normals predicted from RGB (Bethell et al., 2024). "Unified Category-Level Object Detection and Pose Estimation from RGB Images using 3D Prototypes" treats REAL275 as an RGB-only benchmark with known camera intrinsics and emphasizes scale-agnostic metrics to address monocular scale ambiguity (Fischer et al., 4 Aug 2025).

The benchmark has also been repurposed for temporal and relational tasks. CAPTRA evaluates category-level online pose tracking on NOCS-REAL275 from point cloud sequences (Weng et al., 2021). Oryon and Horyon construct cross-scene relative-pose benchmarks from 2,000 REAL275 image pairs, where the same physical object appears in different scenes and must be localized from a textual prompt before registration (Corsetti et al., 2023, Corsetti et al., 2024). Any6D evaluates model-free pose estimation on 2,000 anchor–query RGB-D pairs from REAL275 using only a single RGB-D anchor image per object (Lee et al., 24 Mar 2025), while SingRef6D studies a stricter monocular setting with a single RGB reference and predicted depth (Wang et al., 26 Sep 2025). In a different direction, the universal-features method uses REAL275 as a principal real-world benchmark for zero-shot category-level pose, with leave-one-category-out and reference-mesh protocols for unseen categories (Qu et al., 6 Jan 2025).

3. Evaluation protocols and symmetry conventions

The dominant REAL275 protocol follows NOCS. A prediction is counted as correct when it satisfies either a 3D bounding-box IoU threshold or a joint rotation–translation threshold, and AP or mAP is then computed over categories (Bethell et al., 2024, Zhang et al., 21 Apr 2025). Commonly reported pose thresholds include $5^\circ 2$ cm, $5^\circ 5$ cm, $10^\circ 2$ cm, $10^\circ 5$ cm, and, in some papers, $10^\circ 10$ cm (Lin et al., 2024, Liu et al., 2023). Some works additionally report rotation-only or translation-only thresholds such as $10^\circ$ and 10 cm, or 3D IoU thresholds such as 3D25, 3D50, and 3D75 (Bethell et al., 2024, Liu et al., 2023).

The standard rotation and translation error definitions are explicitly written in several REAL275 papers. Rotation error is given by

$t \in \mathbb{R}^3$0

and translation error by

$t \in \mathbb{R}^3$1

The 3D IoU criterion is defined by the overlap between predicted and ground-truth 3D boxes, for example

$t \in \mathbb{R}^3$2

These formulas appear, with minor notational variations, across NOCS-style evaluations on REAL275 (Bethell et al., 2024, Zhang et al., 21 Apr 2025, Lin et al., 3 Feb 2025).

Symmetry handling is a fixed part of most REAL275 evaluations. For bottle, bowl, and can, rotation around the symmetry axis is ignored; mug is evaluated symmetrically when the handle is not visible (Bethell et al., 2024, Li et al., 20 Aug 2025). Some papers explicitly note this convention, while others state that they follow the standard NOCS protocol; a few summaries note that no special handling for symmetric objects is mentioned in the particular paper even though the broader benchmark convention exists (Zhang et al., 21 Apr 2025).

Derived REAL275 protocols depart from the standard single-image NOCS setup. Oryon and Horyon use BOP-style Average Recall over VSD, MSSD, and MSPD, together with ADD(S)-0.1d, for cross-scene relative pose between two RGB-D views (Corsetti et al., 2023, Corsetti et al., 2024). Any6D also adopts BOP AR, MSSD, MSPD, VSD, and ADD(-S) AUC for its single-anchor, model-free setting on REAL275 pairs (Lee et al., 24 Mar 2025). CAPTRA, by contrast, reports tracking-oriented metrics such as $t \in \mathbb{R}^3$3 cm accuracy, 3D mIoU, rotation error, and translation error over temporal sequences (Weng et al., 2021).

4. Methodological trajectories evaluated on REAL275

REAL275 has supported several distinct methodological families. Early and mid-stage category-level RGB-D work includes prior-based deformation or registration pipelines such as SPD, SGPA, DR-Pose, CATRE, and DPDN. DR-Pose explicitly separates completion-aided deformation from scaled registration, and CATRE treats refinement as iterative alignment between a partial point cloud and an abstract shape prior (Zhou et al., 2023, Liu et al., 2022). DPDN formulates Sim2Real adaptation on REAL275 with a Deep Prior Deformation Network and self-supervised inter- and intra-consistency terms (Lin et al., 2022).

A large set of prior-free or weak-prior RGB-D approaches instead emphasizes direct correspondence learning, implicit canonical mapping, or keypoint learning. IST-Net uses implicit space transformation to move from camera-space features to world-space counterparts without 3D priors (Liu et al., 2023). INKL-Pose predicts 96 instance-adaptive keypoints and refines them with local and global geometric aggregation, including bidirectional Mamba and Feature Sequence Flipping (Zhang et al., 21 Apr 2025). MK-Pose combines RGB, point cloud geometry, and category-level textual descriptions in a multimodal keypoint-learning framework with graph-based relational modeling (Yang et al., 9 Jul 2025). CleanPose adds a causal front-door adjustment module and residual knowledge distillation from ULIP-2, while MemPose introduces an external category-specific geometric memory buffer that stores and updates structural representations from previously observed instances (Lin et al., 3 Feb 2025, Lin et al., 6 Jul 2026).

Representation-learning and modality-transfer approaches are also prominent on REAL275. CLIPose aligns point cloud, image, and text features by multi-modal contrastive learning and prompt tuning, while HRC-Pose learns rotation-aware and translation-aware point-cloud embeddings whose distances track pose continuity (Lin et al., 2024, Li et al., 20 Aug 2025). DeMo-Pose fuses monocular semantic features with depth-based graph features and introduces a Mesh-Point Loss for RGB-D category-level 9-DoF estimation (Agarwal et al., 29 Mar 2026).

REAL275 has additionally become a testbed for RGB-only and open-vocabulary methods. The diffusion-based RGB-only method uses a conditional score-based diffusion model to generate pose hypotheses from a single RGB image, then aggregates them with Mean Shift rather than an additional energy network (Bethell et al., 2024). The unified RGB-only prototype method integrates detection and pose estimation in a single framework by leveraging neural mesh models, dense 2D–3D correspondences, and multi-model RANSAC-PnP (Fischer et al., 4 Aug 2025). In open-vocabulary settings, Oryon and Horyon use text-conditioned localization plus RGB-D registration across paired scenes (Corsetti et al., 2023, Corsetti et al., 2024). Any6D and SingRef6D extend REAL275 to model-free or single-reference novel-object pose, using respectively a single RGB-D anchor or a single RGB reference with predicted depth (Lee et al., 24 Mar 2025, Wang et al., 26 Sep 2025).

This methodological range suggests that REAL275 is not merely a fixed benchmark for one algorithmic template. It functions as a common substrate on which canonical-space regression, deformation, keypoint discovery, diffusion hypothesis generation, contrastive representation learning, memory augmentation, open-vocabulary prompting, and zero-shot reference-based alignment can all be compared under real-scene clutter and category-level generalization.

5. Quantitative landmarks on the benchmark

Because REAL275 is used under several incompatible protocols, its headline numbers are best read within each task family rather than across all papers at once.

Setting	Representative REAL275 result	Paper
RGB-only single-image category-level pose	IoU50 = 45.2; IoU75 = 7.2; 10°/10 cm = 36.6	(Bethell et al., 2024)
Depth-only category-level pose	IoU50 = 83.4; IoU75 = 77.8; 10°5cm = 85.4	(Li et al., 20 Aug 2025)
Depth-only vision–language pose	5°2cm = 48.7%; 10°10cm = 86.2%; 40.0 FPS	(Lin et al., 2024)
RGB-D keypoint-based pose	IoU75 = 80.6; 5°2 cm = 63.2; 10°5 cm = 88.0	(Zhang et al., 21 Apr 2025)
Open-vocabulary relative pose	AR = 57.9; ADD(S)-0.1d = 51.6; mIoU = 81.3	(Corsetti et al., 2024)
Model-free single-anchor relative pose	ADD(-S) AUC = 53.5; AR = 51.0	(Lee et al., 24 Mar 2025)

Within RGB-only single-image category-level pose, the diffusion method reports that its REAL275 configuration with Mean Shift and $t \in \mathbb{R}^3$4 hypotheses reaches IoU50 = 45.2, IoU75 = 7.2, $t \in \mathbb{R}^3$5 = 67.5, 10 cm = 56.6, and $t \in \mathbb{R}^3$6 cm = 36.6, improving over DMSR’s 23.6 at $t \in \mathbb{R}^3$7 cm and 28.3 at IoU50 (Bethell et al., 2024). The same paper states a +55% relative improvement over DMSR on the strict $t \in \mathbb{R}^3$8 cm criterion and a +60% relative improvement on IoU50 (Bethell et al., 2024). In a different RGB-only formulation, the unified prototype method reports 75.2 on NIoU25, 53.7 on NIoU50, 19.2 on NIoU75, and a 22.9% relative average improvement over LaPose across eleven scale-agnostic metrics (Fischer et al., 4 Aug 2025).

Among depth-only or depth-driven category-level methods, HRC-Pose reports IoU50 = 83.4, IoU75 = 77.8, $t \in \mathbb{R}^3$9 cm = 49.8, $s$0 cm = 58.6, $s$1 cm = 72.5, and $s$2 cm = 85.4, surpassing HS-Pose across all listed REAL275 metrics while running at 122.6 FPS (Li et al., 20 Aug 2025). CLIPose reports 48.7% at $s$3 cm, 58.3% at $s$4 cm, 70.4% at $s$5 cm, 85.2% at $s$6 cm, and 86.2% at $s$7 cm, with 40.0 FPS inference (Lin et al., 2024).

Within RGB-D category-level pose, INKL-Pose reports IoU50 = 84.0, IoU75 = 80.6, $s$8 cm = 63.2, $s$9 cm = 68.3, $N_p = 1024$0 cm = 80.9, and $N_p = 1024$1 cm = 88.0 (Zhang et al., 21 Apr 2025). MK-Pose reports IoU50 = 84.0%, IoU75 = 80.3%, $N_p = 1024$2 cm AP = 60.8%, $N_p = 1024$3 cm AP = 78.0%, and $N_p = 1024$4 cm AP = 84.6% (Yang et al., 9 Jul 2025). MemPose reports IoU75 = 81.0, $N_p = 1024$5 cm = 59.9, $N_p = 1024$6 cm = 67.7, and $N_p = 1024$7 cm = 79.0 (Lin et al., 6 Jul 2026). CleanPose, under CATRE-style IoU75*, reports IoU75* = 62.6, $N_p = 1024$8 cm = 61.5, $N_p = 1024$9 cm = 67.4, $5^\circ 2$0 cm = 78.3, and $5^\circ 2$1 cm = 86.2, improving over AG-Pose by +4.5 points at $5^\circ 2$2 cm (Lin et al., 3 Feb 2025).

REAL275 has also been central in relative-pose and novel-object settings. Horyon reports AR = 57.9 and ADD(S)-0.1d = 51.6 on 2,000 REAL275 cross-scene pairs, exceeding Oryon by +25.5 AR and +27.3 ADD in that open-vocabulary RGB-D setup (Corsetti et al., 2024). Any6D reports ADD(-S) AUC = 53.5, AR = 51.0, MSSD = 56.5, MSPD = 65.3, and VSD = 31.1, outperforming Oryon on most metrics except VSD (Lee et al., 24 Mar 2025). SingRef6D, using a single RGB reference and predicted depth, reports AR = 28.7 on REAL275 and a +14.41 percentage-point improvement in $5^\circ 2$3 over Depth-Anything v2 for depth prediction on REAL275 (Wang et al., 26 Sep 2025).

6. Recurrent difficulties, ambiguities, and benchmark significance

Several failure modes recur across REAL275 studies. Scale ambiguity remains a persistent issue whenever methods avoid metric depth or instance CAD models. The diffusion-based RGB-only paper states that scale ambiguity persists in challenging cases and that translation mAP is sensitive to context and global features (Bethell et al., 2024). The unified RGB-only prototype method evaluates both scale-agnostic and scale-aware metrics precisely because single-view RGB introduces a fundamental ambiguity between object distance and metric scale (Fischer et al., 4 Aug 2025). MSL-RAPTOR, which uses only monocular RGB, identifies depth along the optical axis as the main source of translation error on REAL275 (Ramtoula et al., 2020).

Symmetry is another structural difficulty. Rotation around symmetry axes is underdetermined for bottles, bowls, and cans, and mug evaluation changes when the handle is not visible (Bethell et al., 2024, Li et al., 20 Aug 2025). Some methods explicitly model this during training, such as MK-Pose’s symmetry-aware pose-size loss (Yang et al., 9 Jul 2025), while open-vocabulary and BOP-style protocols absorb it into symmetry-aware metrics such as MSSD, MSPD, and ADD-S (Corsetti et al., 2023, Corsetti et al., 2024). A plausible implication is that REAL275 measures not only raw rotation recovery but also the ability to represent or tolerate pose equivalence classes induced by object geometry.

Intra-class variation and sensor quality also matter strongly. The diffusion paper remarks that camera and mug categories show lower rotation mAP because different shapes, lenses, and handle visibility lead to distribution spread and estimation uncertainty (Bethell et al., 2024). CAPTRA reports that the camera category is especially difficult because severe real depth noise can break depth-only segmentation (Weng et al., 2021). DR-Pose notes that mug handles are often occluded or invisible in REAL275, and that completion can mis-hallucinate them (Zhou et al., 2023).

Derived REAL275 protocols reveal additional sensitivities. In open-vocabulary relative pose, Oryon shows that segmentation quality, camera distance between views, and prompt descriptiveness materially affect AR, and that generic or misleading prompts can sharply reduce performance (Corsetti et al., 2023). SingRef6D notes failure cases in extremely dark scenes where RGB lacks information, while also showing that better monocular depth substantially improves downstream registration on REAL275 (Wang et al., 26 Sep 2025).

Taken together, these observations suggest that REAL275 remains influential because it concentrates several hard factors in one real benchmark: cluttered indoor scenes, occlusion, real sensor artifacts, novel-instance generalization, symmetry, and scale ambiguity. Its continued reuse across RGB-D, depth-only, RGB-only, open-vocabulary, tracking, single-anchor, and zero-shot formulations indicates that it functions as both a benchmark dataset and a benchmark design pattern for studying how category-level geometric reasoning transfers from controlled canonical spaces to real-world visual scenes (Bethell et al., 2024, Li et al., 20 Aug 2025, Corsetti et al., 2024).