Object Pose Estimation · 2025 · CVPR · CCF A

One2Any: One-Reference 6D Pose Estimation for Any Object

One2Any encodes a single reference RGB-D image into ROPE and decodes Reference Object Coordinates for a query, replacing multiview retrieval and CAD alignment with a continuous relative-coordinate representation.

Paper Code

Research area

Object Pose Estimation

Task level

Instance-level

Object setting

Unseen objects

Input modalities

single reference RGB-Dquery RGB-D

Requirements

reference maskquery maskcamera intrinsics

Full paper metadata Authors, affiliations, research area, datasets, metrics, method tags, and reading status Expand

Authors: Mengya Liu, Siyuan Li, Ajad Chhatkuli, Prune Truong, Luc Van Gool, Federico Tombari
Affiliations: ETH Zurich / INSAIT / Sofia University / Google / Technical University of Munich
Conference / journal: Conference · CVPR · CCF A
Research area: Object Pose Estimation
Code availability: Code available
Reading status: Overview:Available Deep Dive:Available
Task setting: Instance-level · Unseen objects
Datasets: REAL275 Toyota-Light YCB-Video LINEMOD
Evaluation metrics: ADD / ADD-S AUC ADD-S ADD BOP AR VSD MSSD MSPD
Method tags: Novel-object Pose Estimation Relative Pose Estimation 2D–3D Correspondences Dense Correspondences & Coordinate Fields Synthetic Data & Foundation Models
Requirements: reference maskquery maskcamera intrinsics

Reading Notes

Overview / Deep Dive

Available src/paper-notes/one2any/coarse.md

Paper overview notes

Quickly assess the paper's scope, problem setting, core method, experimental findings, and whether to read it in depth.

Paper Reading Notes (Coarse): One2Any: One-Reference 6D Pose Estimation for Any Object

1. Basic information

Paper: CVF Open Access
Authors / affiliations: Mengya Liu et al.; ETH Zurich, INSAIT, Google, TUM.
Task: relative 6D pose for arbitrary objects from one reference RGB-D and one query RGB-D image.
Reading mode: Overview.

2. Positioning in one sentence

One2Any compresses reference shape, pose, and texture into ROPE, decodes query ROC maps without canonical NOCS, and solves relative pose with Umeyama.

3. Problem and motivation

CAD, many references, or category-specific training raise onboarding costs.
Single-reference estimation must preserve the reference frame, beyond matching features.

4. Method overview

A Reference Object Encoder produces ROPE; pretrained VQ-VAE encodes queries.
Cross-attentive U-Net fuses ROPE / query features to output Reference Object Coordinates.
Align ROC with query depth through Umeyama for relative rigid transforms.

5. Experiments at a glance

REAL275, Toyota-Light, occluded YCB-Video, LINEMOD, and tracking compare one-reference with multi-view / CAD approaches.
Main tables report reference-relative pose, distinct from canonical absolute-pose outputs.

One2Any single-reference REAL275 and Toyota-Light results — Original REAL275 / Toyota-Light tables: 2,000 reference-query pairs per test set. One2Any AR / ADD-0.1d is 54.9 / 41.0 and 42.0 / 34.6, respectively. Source: CVPR 2025 paper. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

Single-reference RGB-D Oryon AR is 46.5 / 34.1 on REAL275 / Toyota-Light versus One2Any 54.9 / 42.0; ADD-0.1d rises from 34.9 / 22.9 to 41.0 / 34.6. This is more direct than cross-modality comparisons but still assumes GT masks and the prescribed 2,000 pairs.

6. Quick assessment

Value: reference-bound ROC supervision avoids canonicalizing unknown objects.
Risks: occluded references, depth holes, and absent query-reference overlap directly limit solvability.

Available src/paper-notes/one2any/deep.md

In-depth paper notes

Support reproduction, comparison, related-work writing, and identifying opportunities for further research.

Paper Reading Notes (Deep): One2Any: One-Reference 6D Pose Estimation for Any Object

1. Key assessment

Paper: CVF Open Access
Question: without CAD or canonical NOCS and with only one reference, what coordinate supervision can large paired datasets teach?
Answer: use the reference-camera frame as the sole anchor and predict ROC. ROPE conditions query decoding on reference texture, shape, and pose.

2. Positioning in one sentence

One2Any continuously maps one-reference object embeddings to reference-coordinate query decoding, then Umeyama rigid fitting for relative pose.

2.1 Guide to the paper’s figures and tables

Figure / table	What to look for
Figure 2	ROE, ROPE, VQ-VAE, cross-attentive U-Net, and ROC information paths
Figure 3	ROC versus NOCS: reference frames avoid unknown-category canonicalization
Tables 1–4	Different metrics for REAL275 / Toyota-Light, occluded YCB-Video, large-view LINEMOD, and tracking
Table 5	Speed alongside CAD / multi-view reference budgets

One2Any original ROPE / ROC single-reference pose network — Original Figure 2: Reference Object Pose Embedding encodes one RGB-D reference; query decoding predicts Reference Object Coordinates, followed by Umeyama pose fitting. Source: CVF Open Access. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

3. Problem setting and boundaries

Both reference / query are RGB-D; no external reference-object pose annotation.
Output is the query-relative-to-reference rigid transform, with different frame freedoms from canonical CAD pose.
Masks and query-depth fitting are required; this is not RGB-only single-reference inference.

4. Method details

ROE / ROPE: encode reference shape, appearance, and orientation as latent conditions.
Query decoding: pretrained VQ-VAE features enter a ROPE-conditioned cross-attentive U-Net, predicting per-pixel reference-frame ROC.
Pose solving: ROC / depth-back-projected query pairs enter Umeyama for relative rotation, translation, and scale relationships.

5. Experimental analysis

REAL275, Toyota-Light, occluded YCB-Video, large-view LINEMOD, and tracking test categories, occlusion, and sequential use.
Group by reference conditions; one RGB-D reference cannot be declared unconditionally better than 8–200 views or CAD.
Embedding / decoding avoids per-query multi-view search or optimization for speed, depending on learned object-latent generalization.

One2Any occluded YCB-Video ADD-S / ADD AUC results — Original occluded YCB-Video results mix 16 views, CAD, single CAD, and single RGB-D references. Group reference budgets before comparing ADD-S / ADD. Source: CVPR 2025 paper. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

One2Any’s one-reference occluded YCB-Video mean ADD-S / ADD AUC is 93.7 / 84.4; Oryon gives 13.3 / 7.4 and NOPE with GT translation 86.0 / 25.1. Single-CAD FoundationPose gives 90.4 / 76.1, versus 97.4 / 91.5 with CAD generated from 16 views. This supports competitive occluded single-reference estimation while acknowledging greater information in complete 16-view CAD.

One2Any large-view LINEMOD results and YCB-Video qualitative comparisons — Original large-view LINEMOD table, YCB-Video / LINEMOD qualitative comparisons, and tracking results. Green predictions and pink GT are shown, with ROC maps at the right. Source: CVPR 2025 paper. Image copyright belongs to the paper's authors and publisher; reproduced here for academic analysis.

Large-view LINEMOD ADD-0.1d mean recall is 52.6, above one-reference FoundationPose 48.3, NOPE 16.3, and Oryon 9.8, below 16-RGB-D FS6D+ICP 91.5. Eggbox / glue reach 100.0 / 99.9, versus ape / benchwise 33.1 / 15.7, reflecting object-dependent symmetry, texture, and reference coverage. Fixed-first-frame YCB-Video tracking gives ADD-S / ADD 93.8 / 84.8; single-CAD FoundationPose 80.9 / 57.9 and 16-CAD 97.5 / 93.7.

Reported RTX 4090 runtime is 0.09 s, versus Oryon 0.90 s and FoundationPose 2.70 s. Feedforward latency benefits come with unequal GPU / rendering / onboarding conditions. ROC versus direct rotation-translation heads raises ADD-S / ADD from 84.7 / 59.4 to 91.2 / 78.6 and 5°5cm from 2.4 to 15.5, supporting correspondence decoding for generalization.

6. Authors’ claims and my assessment

Authors’ claims	My assessment
One reference rivals multi-view / CAD	Competitive across benchmarks, but check shared masks, depth, and protocols per row
ROC suits unknown objects better than NOCS	Reference frames avoid unknown canonicalization, while every query remains relative to that reference

7. Reproduction considerations and limitations

Check reference augmentation, ROC normalization, masks, and Umeyama scale handling first.
Occluded / narrow references, no overlap, and inaccurate depth remove correspondence information, beyond a small correction problem.
Absolute world pose requires additional reference-to-robot / scene registration.
Training uses FoundationPose synthetic pairs and OO3D-9D multi-view data. Separate training object distributions from test instances; unseen instances do not imply independence from large-scale object priors.

Related papers

2026 · CVPR

3PT

3PT unifies detection, segmentation, and 6-DoF pose with two Transformers for 2D / 3D object-centric understanding. Continuous refinement of object representations and rendered-feature alignment support zero-shot cross-object perception on BOP-H3, BOP-Industrial, and BOP-Classic.

2026 · AAAI

6DAttack

Studies backdoor attacks on PnP-based and end-to-end 6-DoF pose pipelines. Structurally 3D triggers and modified pose labels are injected into training data so that models behave normally on clean inputs but produce controlled incorrect poses when triggered.

2026 · ICRA

Active Tactile Exploration

A violation-implicit loss incorporates physical-constraint violations into joint estimation of freely moving unknown convex rigid-body trajectories and cuboid / 20-vertex polytope shapes. Observed / Fisher information at a local MLE estimates EIG to select the next touch. In simulation, the EIG strategy yields stable errors below 4 mm after 5 cuboid actions; real Trifinger results at step 5 substantially outperform random selection. Convex geometry, local optima, multiple initializations, and tactile / proprioceptive errors limit the method.