acceptodds
Under review as a conference paper at ICLR 2027

GROUND2POSE: GEOMETRIC SELF-VERIFICATION FOR TRAINING-FREE OPEN-VOCABULARY 6D POSE ESTIMATION

Abstract

Estimating the 6D pose of novel objects is critical for open-world embodied perception, yet existing estimators rely heavily on object CAD models, multi-view calibration, or task-specific fine-tuning. While emerging language-guided methods bypass pre-scanned 3D assets, they tightly couple 2D localization with 3D pose recovery, allowing upstream segmentation errors to catastrophically corrupt downstream geometric reasoning. We present Ground2Pose, a training-free framework that breaks these constraints by estimating 6D relative object pose from an anchor–query RGB-D view pair and a text description, without CAD assets, reference calibration, or task supervision. Ground2Pose explicitly decouples open-vocabulary 2D localization from 3D geometric reasoning. At its core is the Geometric Self-Verification Solver (GVS)—a solver that turns its own closed-form alignment into an outlier verifier, systematically pruning correspondence outliers via 3D alignment residuals rather than unreliable semantic or heuristic confidence scores. On four diverse benchmarks under our strict single-anchor, predicted-mask protocol, Ground2Pose achieves SOTA performance (70.9 ADD(S)-0.1d and 42.2 BOP recall), substantially outperforming the strongest comparable competitor by +40.8 and +8.8 points. Ablations and controlled mask interventions reveal that while upstream localization sets the empirical accuracy ceiling, correspondence-level geometric verification determines how closely the solver approaches it, establishing a robust verification-over-reweighting design principle. With fixed pre-trained backbones and closed-form efficiency, Ground2Pose provides a reproducible and robust foundation for open-world spatial grounding. Code will be made publicly available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
6D Pose Estimation, Training-Free, Language-Guided Localization, Geometric Self-Verification, Open-Vocabulary Segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_y3kQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Ground2Pose, a training-free framework for generalizable 6D object pose estimation. It uses a single source–reference RGBD image pair and a text description, without requiring CAD models or task-specific training. The framework consists of language-guided object localization, pixel matching, and relative pose estimation. Experiments on four benchmarks show promising results, outperforming the other evaluated competitors.

Critical strengths
  • The proposed method supports generalizable 6D object pose estimation without requiring CAD models, multiple reference views, or task-specific training, making it applicable to novel objects.
  • The authors evaluate the method across four benchmarks, comparing it against several representative baselines. The presented approach achieves better performance, demonstrating the effectiveness of the training-free framework.
  • The ablation studies show consistent improvements over standard rigid alignment and alternative robust estimators such as RANSAC, TEASER++, and SC2-PCR.
Critical weaknesses
  1. The technical contribution of this paper is a bit limited and incremental. The overall pipeline follows a relatively straightforward design: using an existing open-vocabulary object detector to detect the novel object and handling the subsequent object pose estimation via a point cloud registration module. Such a design is quite straightforward and has been used in previous work. For instance, [1] also uses an open-vocabulary object detector for object detection. Besides, the authors use LoFTR to build correspondences, which can be replaced with other advanced approaches, such as roma [2] or vggt-omega [3].
  2. The reliance on RGBD input significantly simplifies the object pose estimation problem, as depth information enables direct 3D correspondence construction and rigid registration. However, this assumption may limit its applicability in some scenarios where depth is noisy or unavailable.
  3. The authors list some powerful relevant methods in Table 1, such as FoundationPose, but a comparison with it is missing.

Overall, I am concerned that the technical novelty and contributions of this work may not meet the standards expected for publication at ICLR.

[1] Chen Zhao, Tong Zhang, Zheng Dang, and Mathieu Salzmann. "Pose without Guesses: Generalizable Object Pose Estimation From a Single Reference." International Journal of Computer Vision 134, no. 3 (2026): 136.

[2] Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. "Roma: Robust dense feature matching." In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 19790-19800. IEEE, 2024.

[3] Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. "VGGT-Ω\Omega." arXiv preprint arXiv:2605.15195 (2026).

Questions for authors

Please refer to my concerns in Weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer y3kQ (Part 1/2)

We thank Reviewer y3kQ for recognizing the practical value of our CAD-free, single-anchor, training-free paradigm for novel objects and our consistent improvements across four benchmarks and multiple robust estimators. Below we address each concern in detail.

1. Technical Contribution, Architectural Insight, and New Literature (W1, W4)

[W1 & W4]: The technical contribution of this paper is a bit limited and incremental... [1] also uses an open-vocabulary object detector... LoFTR can be replaced with other advanced approaches, such as roma [2] or vggt-omega [3].

We thank the reviewer for pointing out DVMNet++ ("Pose without Guesses", IJCV 2026) [1], RoMa (CVPR 2024) [2], and VGGT-Ω (2026) [3], all of which are now cited and discussed in Section 2, Section 3.2, and Appendix A/K. Three key points distinguish our contribution:

  1. Architectural Decoupling vs. Trained Latent Alignment ([1], Oryon, Horyon, One2Any): While DVMNet++ [1] and One2Any (CVPR 2025) also address single-reference pose estimation without CAD models, they are task-trained neural networks that lift reference/query views into learned voxelized 3D latents ([1]) or regress dense object coordinates (One2Any, requiring ~2,700 GPU-hours of synthetic training). Similarly, open-vocabulary methods like Oryon and Horyon train multimodal fusion backbones. A prevalent assumption in these works is that end-to-end learned fusion or semantic-confidence weighting is necessary to bridge noisy 2D detections and 3D pose. Our paper challenges this paradigm: we prove both constructively and through 8 pre-registered negative experiments (Appendix C.2 & H.6) that coupling or soft-reweighting 2D semantics with 3D geometry actually degrades pose accuracy (by up to −10.0 points), whereas a strictly decoupled, training-free pipeline with hard 3D correspondence self-verification (GVS) outperforms trained single-reference SOTA (beating One2Any by +8.78 points average and +21.60 points on cluttered NOCS, and beating Oryon/Horyon by +43.5 to +54.7 points).
  2. Giving Existing Methods the Same Off-the-Shelf Detector Does Not Close the Gap: In Table 2 (Prior = Ours), all baselines (Oryon, LatentFusion, OnePoseviaGen, GigaPose, and One2Any in Table 15) are evaluated using our exact same Grounded-SAM masks, and even under 100% Ground-Truth Oracle masks (Prior = Oracle), they still trail Ground2Pose by 17.0 to 63.2 points. Thus, the performance leap comes from our match-then-self-verify formulation, not merely from using an off-the-shelf detector.
  3. New Experiments Replacing LoFTR with RoMa [2] (Appendix K.3): Following your suggestion, we replaced LoFTR with RoMa [2] (roma_outdoor.pth) with zero changes to GVS (γ=0.7\gamma=0.7). As detailed in our Global Response (§1) and Response to JCkW (§2), RoMa + GVS achieves 40.7% / 77.0% / 30.0% on YCBV / NOCS / LM, still outperforming all prior Table 2 baselines, and GVS boosts RoMa across all four datasets—including a massive +17.80-point gain on wide-baseline NOCS (59.16% → 76.96%) where RoMa's dense warp field produces silhouette outliers that corrupt plain Kabsch. We also discuss feed-forward 3D foundation models like VGGT-Ω [3] in Section 3.2 as promising correspondence front-ends that naturally plug into our GVS backend.

2. Positioning Relative to FoundationPose (W3)

[W3]: The authors list some powerful relevant methods in Table 1, such as FoundationPose, but a comparison with it is missing.

As shown in Table 1 (tab:positioning), FoundationPose operates under a fundamentally different assumption regime:

  • Its model-based mode requires a pre-scanned 3D CAD mesh of the target object at inference time;
  • Its model-free mode requires a multi-view RGB-D reference sequence (~16+ views covering the object) to optimize a per-object neural implicit field prior to testing, and it does not perform open-vocabulary text grounding.

In contrast, our protocol provides only a single anchor RGB-D frame and a text prompt, with zero CAD models and zero per-object optimization. Note that in Table 2 we already evaluate GigaPose (which receives 3D CAD templates) and LatentFusion / OnePoseviaGen (which build neural/generative multi-view representations from the single anchor) to represent these paradigms under the single-anchor constraint. We have expanded Section 2 and Appendix A to make this regime boundary with FoundationPose explicit.

AuthorsOpenReview ↗

Response to Reviewer y3kQ (Part 2/2)

3. Reliance on RGB-D Input, Sensor Depth Noise, and New RGB-Only Query Extension (GVS-PnP) (W2)

[W2]: The reliance on RGBD input significantly simplifies the object pose estimation problem... However, this assumption may limit its applicability in some scenarios where depth is noisy or unavailable.

We address both noisy depth and unavailable query depth with new quantitative experiments (Appendix K.2 & K.4):

  1. Why RGB-D is standard for CAD-free metric 6D pose: Without a 3D CAD model of known physical size or category-specific scale supervision, recovering metric 6D translation [R∣t][\mathbf{R}|\mathbf{t}] from a single reference view is inherently scale-ambiguous unless metric depth is observed at least in the anchor frame (the exact same RGB-D modality used by Oryon, Horyon, LatentFusion, One2Any, and FoundationPose in Table 1).
  2. Robustness to Noisy Sensor Depth (Appendix K.2): To evaluate noisy RGB-D sensors, we injected multiplicative Gaussian depth noise σz∈{0.5,1.0,2.0}\sigma_z \in \{0.5, 1.0, 2.0\} into both anchor and query depth maps before 3D back-projection. Even at heavy 2.0% depth noise (exceeding commercial RealSense/Kinect sensor error at manipulation distances), ADD(S)-0.1d changes by only −2.04 (YCBV), −1.65 (NOCS), −2.44 (TOYL), and +0.96 (LM) (mean shift −1.29 points), demonstrating that GVS's residual trimming effectively absorbs sensor depth noise.
  3. When Query Depth is Completely Unavailable — Preliminary RGB-Only Query Extension (GVS-PnP, Appendix K.4): What if the query observation is RGB-only (no query sensor depth)? While naively feeding monocular predicted depth (e.g., ZoeDepth or Depth-Anything-V2) into 3D–3D SVD fails due to global monocular scale/affine drift (0.0%–7.4%), our GVS principle naturally extends to RGB-only queries by replacing 3D–3D SVD with 2D–3D PnP + 2D Reprojection Self-Verification (GVS-PnP) using only anchor depth:
Query Modality & MethodYCBVNOCSTOYLLM
Oryon (Full RGB-D Anchor & Query, Trained)11.4023.3022.907.20
LatentFusion (Full RGB-D, Trained)24.4018.7054.6022.80
Ground2Pose w/ GVS-PnP (RGB-Only Query, No Query Depth)31.0035.4027.4015.95
Ground2Pose w/ 3D–3D GVS (Full RGB-D, Default)58.5086.2083.2055.60

Remarkably, even without any query depth map, our training-free GVS-PnP extension achieves 31.00% on YCBV and 35.40% on NOCS, outperforming full RGB-D baselines such as Oryon (11.4% / 23.3%) and LatentFusion (24.4% / 18.7%)! This confirms that our decoupled self-verification paradigm is not restricted to 3D–3D RGB-D registration and extends cleanly to RGB-only query scenarios.

We hope these new experiments, RoMa ablations, and RGB-only query results address your concerns.

I thank the authors for their response. However, my concerns regarding the technical novelty and advantages over FoundationPose still exist.

The idea is straightforward, and the novelty is a bit limited. As I mentioned in my reviews, some open-vocabulary object detectors exist and work well. Besides, with the availability of powerful foundation models such as SAM2, open-vocabulary object detection does not seem to be a particularly challenging problem.

A direct comparison with FoundationPose is still missing. The proposed method also requires RGB-D inputs, similar to FoundationPose, with the key distinction being its reliance on a single reference view rather than multiple views. In this context, it is important to quantify how much performance is sacrificed in exchange for reducing the dependence on prior observations. If the performance degradation is substantial, it remains unclear whether the benefits of the proposed setup outweigh this trade-off.

Therefore, I have to keep my original rating.

Reviewer_JCkWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Ground2Pose, a training-free framework for open-vocabulary 6D object pose estimation using an anchor–query RGB-D image pair and a natural-language description, without requiring CAD models or task-specific training. The framework first uses Grounded-SAM to localize the target object, then applies LoFTR to establish cross-view correspondences within the segmented regions. These correspondences are lifted into 3D using depth information, and an initial rigid transformation is estimated via Kabsch alignment. The proposed GVS further refines the pose by filtering correspondences based on geometric residuals and re-estimating the transformation. The method is evaluated on four benchmarks, with experiments examining pose estimation accuracy, localization quality, and geometric solver robustness.

Critical strengths
  1. Clear and practical framework. The paper addresses a relevant setting for open-vocabulary 6D object pose estimation without CAD models or task-specific training. The proposed pipeline is straightforward, clearly presented, and easy to understand.

  2. Strong empirical results and extensive evaluation. The paper reports substantial improvements across four benchmarks. The supplementary material includes comparisons with alternative localization modules and geometric solvers, hyperparameter studies, per-object results, and statistical analyses, providing useful empirical evidence for the effectiveness of the overall pipeline.

Critical weaknesses
  1. Incomplete coverage of closely related work and baselines. The related work and experimental comparisons overlook several highly relevant studies, such as One2Any (CVPR 2025), which addresses single-reference relative 6D pose estimation, and FiCoP (2026), which focuses on robust cross-view correspondence for open-vocabulary 6D pose estimation. Furthermore, closely related methods such as Oryon and Horyon already employ robust geometric registration, yet PointDSC is not included in the controlled solver comparison.

  2. Limited methodological novelty and insufficient attribution of performance gains. The proposed framework primarily combines existing components, including Grounded-SAM, LoFTR, and Kabsch alignment, while GVS closely resembles conventional residual-based trimmed least squares. Although the paper reports substantial improvements over existing methods, most of the performance advantage is already achieved by the basic pipeline without GVS (67.3% vs. 70.9% average ADD(S)-0.1d). Further diagnostic experiments are needed to investigate whether applying the same off-the-shelf components to existing methods could yield comparable performance improvements.

[1] One2Any: One-Reference 6D Pose Estimation for Any Object. CVPR 2025. [2] FiCoP: Learning Fine-Grained Correspondence with Cross-Perspective Perception for Open-Vocabulary 6D Object Pose Estimation. 2026.

Questions for authors
  1. Module interchangeability. The framework is described as modular, but the correspondence matching backbone remains fixed to LoFTR. How sensitive is the performance to different off-the-shelf matching models? Could similar components also improve existing approaches?

  2. Anchor–query pair difficulty. How are the anchor–query pairs sampled? Could the authors provide a breakdown of performance across different viewpoint changes and overlap ratios?

  3. Fairness of solver comparisons. In Appendix H.4, the thresholds for different robust solvers are derived from the initial Kabsch residuals. How does this choice affect the comparison, and would individually tuned thresholds change the conclusions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer JCkW (Part 1/3)

We thank Reviewer JCkW for appreciating our clear, practical training-free framework and extensive empirical and statistical evaluation. We have conducted all requested comparisons (One2Any, PointDSC, RoMa matcher swap, and viewpoint stratification) and updated Section 2, Table 4, and Appendix A/H/K.

1. New Baselines & Related Work: One2Any (CVPR 2025), FiCoP (2026), and PointDSC (W1)

[W1]: The related work and experimental comparisons overlook several highly relevant studies, such as One2Any (CVPR 2025)... and FiCoP (2026)... PointDSC is not included in the controlled solver comparison.

  • Quantitative Comparison with One2Any (Appendix K.1, Table 15): We evaluated One2Any on our exact 2,000-pair manifests across all four benchmarks using both our Grounded-SAM masks (Ours) and GT masks (Oracle). As shown in the Global Response (§3), training-free Ground2Pose outperforms One2Any on all four benchmarks under both mask settings:
    • Predicted Masks (Ours): Ground2Pose achieves 58.50 / 86.20 / 83.20 / 55.60 (Avg: 70.88%) vs. One2Any's 50.65 / 64.60 / 81.75 / 51.40 (Avg: 62.10%)—leading by +7.85 (YCBV), +21.60 (NOCS), +1.45 (TOYL), and +4.20 (LM).
    • Oracle GT Masks: Ground2Pose leads by +16.99 points on average (87.70% vs. 70.71%).
    • Structural Insight: One2Any requires ~2,700 GPU-hours of synthetic training (8× RTX 4090 for two weeks) to learn direct object-coordinate regression from pre-segmented masks. While competitive on clean single-object scenes (TOYL, +1.45 gap), its coordinate regression degrades under heavy clutter, occlusion, and cross-instance ambiguity (NOCS, +21.60 gap), where Ground2Pose's correspondence-level geometric self-verification excels without any training.
  • Controlled Solver Comparison with PointDSC (Table 4 & Appendix H.4): We evaluated PointDSC (official 3DMatch checkpoint) on the exact same LoFTR correspondences and γ=0.7\gamma=0.7 aperture across all 8,000 pairs (Global Response §2). PointDSC reaches 55.25 / 83.20 / 79.95 / 52.95 (Avg: 67.84%)—outperforming SC2-PCR (60.29%) and edging out RANSAC-256 on NOCS (83.20% vs. 82.60%), yet trailing GVS (58.50 / 86.20 / 83.20 / 55.60, Avg: 70.88%) on every benchmark by 2.65–3.25 points, while GVS is 30× faster (0.16 ms vs. 4.8 ms).
  • Related Work Coverage: One2Any (CVPR 2025), FiCoP (2026), and PointDSC (CVPR 2021) are now thoroughly discussed in Section 2, Table 1, and Appendix A.
AuthorsOpenReview ↗

Response to Reviewer JCkW (Part 2/3)

2. Attribution of Gains & Matcher Interchangeability (W2, Q1)

[W2 & Q1]: ...most of the performance advantage is already achieved by the basic pipeline without GVS (67.3% vs. 70.9%). Further diagnostic experiments are needed to investigate whether applying the same off-the-shelf components to existing methods could yield comparable performance improvements... How sensitive is the performance to different off-the-shelf matching models?

  • 1) Applying the exact same off-the-shelf localization to existing methods was already tested in Table 2: In Table 2 (Prior = Ours), all evaluated baselines (Oryon, LatentFusion, OnePoseviaGen, GigaPose) and One2Any (Table 15) receive the exact same Grounded-SAM masks as Ground2Pose. Even under 100% perfect GT masks (Prior = Oracle), Oryon achieves only 24.5%, GigaPose 25.8%, LatentFusion 30.8%, and One2Any 70.7%—all far below Ground2Pose (87.7%). Because prior pipelines entangle segmentation boundaries with feature fusion or coordinate regression, simply giving them Grounded-SAM does not close the gap. Our primary contribution is demonstrating that strictly decoupling 2D open-vocabulary localization from 3D correspondence verification (and avoiding soft semantic reweighting, validated by 8 negative probes in Appendix C.2 & H.6) unlocks this performance leap without training.
  • 2) Why GVS is essential beyond the +3.5-pt average under default masks: Under default masks, GVS yields +17.5 to +25.8 points on hard cluttered/symmetric objects (Tables 6–9), eliminates 21.7% of remaining errors under GT masks (84.3% → 87.7%), and boosts accuracy by +10.40 to +17.35 points under noisier localization (Talk2DINO, Table 4) and +17.80 points under dense warp matching (RoMa on NOCS, below). Unlike iterative Trimmed ICP—which re-assigns nearest neighbors each pass and requires close pose initialization—GVS operates on fixed putative cross-view matches in a single closed-form pass (Kiter=1K_{\text{iter}}=1, 0.16 ms).
  • 3) Matcher Interchangeability — Swapping LoFTR for RoMa (Appendix K.3): We replaced LoFTR with the off-the-shelf dense warp matcher RoMa (roma_outdoor.pth) on 200-pair probes across all four datasets with GVS parameters untouched (γ=0.7\gamma=0.7):
Matcher ConfigurationYCBVNOCS (Wide Baseline)TOYLLM
RoMa + Plain Kabsch (w/o GVS)39.5659.1620.1127.89
RoMa + GVS (Ours)40.6676.9623.9130.00
GVS Gain on RoMa+1.10+17.80+3.80+2.11
LoFTR + GVS (Default)58.5086.2083.2055.60

Two insights emerge: (i) GVS improves RoMa on every dataset, including a +17.80-point surge on NOCS (59.16% → 76.96%), because RoMa's continuous warp field interpolates across object silhouettes, creating boundary outliers that corrupt plain Kabsch but are cleanly pruned by GVS; (ii) even without matcher-specific tuning, RoMa + GVS beats all prior Table 2 baselines on YCBV, NOCS, and LM. LoFTR remains our default because Stage 2 zeros background pixels prior to matching (I~=I⊙M∗\tilde{I} = I \odot M^*, Eq. 1), where LoFTR's local coarse-to-fine correlation retains higher match density on small texture-poor crops (TOYL/LM) than global ViT patch tokens.

AuthorsOpenReview ↗

Response to Reviewer JCkW (Part 3/3)

3. Anchor–Query Sampling and Viewpoint-Change Breakdown (Q2)

[Q2]: How are the anchor–query pairs sampled? Could the authors provide a breakdown of performance across different viewpoint changes and overlap ratios?

  • Sampling Protocol: Following Oryon/Horyon, for each dataset we sample 2,000 cross-view anchor–query pairs (fixed seed 42) where the target object is visible in both frames; on NOCS-REAL275, anchors and queries originate from distinct camera streams (camera_canon vs. camera_shengjun), enforcing wide cross-perspective baselines by construction.
  • Viewpoint-Change Breakdown (Appendix K.2, Table 16): Stratifying all 8,000 pairs by ground-truth relative rotation θ\theta (full table in our Response to Reviewer 4B8a §4) shows that 75.3% of all pairs (6,022 / 8,000) have large viewpoint changes (> 60°), with median rotation spans of 66.5° (YCBV), 117.0° (NOCS), 112.3° (TOYL), and 96.0° (LM). Accuracy decreases smoothly with viewpoint gap (e.g., YCBV: 71.63% at < 30°, 57.80% at 30°–60°, 51.81% at > 60°; TOYL: 95.45% → 90.99% → 79.61%), and even in the > 60° low-overlap regime, Ground2Pose achieves 51.8% / 86.2% / 79.6% / 51.9%, surpassing all competing baselines' overall averages.

4. Fairness of the Shared Solver Aperture vs. Fixed Tuned Thresholds (Q3)

[Q3]: In Appendix H.4, the thresholds for different robust solvers are derived from the initial Kabsch residuals. How does this choice affect the comparison, and would individually tuned thresholds change the conclusions?

  • Why the shared γ\gamma-quantile aperture actually helps the competing solvers: Across our four benchmarks, object diameters span nearly an order of magnitude (from small LineMOD objects < 10 cm to large YCBV/NOCS containers), and effective residual scales vary per pair with camera distance and viewpoint span. Using a fixed scene-scale metric threshold (e.g., PointDSC's native 10 cm 3DMatch default) causes severe scale mismatch on object-scale point clouds, dropping PointDSC on YCBV from 55.25% (with our shared adaptive aperture) down to 30.15% (with its fixed threshold). Thus, calibrating each solver's inlier/noise bound to the per-pair γ=0.7\gamma=0.7 residual quantile provides automatic per-instance scale normalization for RANSAC, TEASER++, SC2-PCR, and PointDSC.
  • Why individual threshold tuning does not change the ranking: As shown in Appendix H.1 (Table 12), sweeping the aperture across γ∈[0.5,0.9]\gamma \in [0.5, 0.9] shifts accuracy by ≤ 1.1–1.8 points, far smaller than the margins over TEASER++ (−27.9 pts), SC2-PCR (−10.6 pts), and PointDSC (−3.0 pts). The baselines trail because mask boundary leakage creates spatially coherent background clusters that violate max-clique and spatial-compatibility assumptions regardless of scalar threshold tuning.

We hope our new baselines (One2Any, PointDSC), matcher ablations (RoMa), and viewpoint stratifications thoroughly address your concerns.

Thank you for the detailed rebuttal and additional experiments. The new comparisons and analyses address most of my concerns regarding experimental completeness.

However, my main concern about methodological novelty remains. While the results demonstrate the effectiveness of the framework, the overall pipeline largely follows established geometric vision techniques, and GVS closely resembles conventional residual-based trimming. I remain unconvinced that the proposed decoupling and verification strategy constitutes a sufficiently distinct methodological contribution.

Reviewer_4B8aOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Given text, an RGB-D pair, and an anchor view with known pose, the method segments the object with an open-vocabulary model, matches it with LoFTR, and estimates relative 6D pose in closed form. GVS drops correspondences with large alignment residuals and solves once more. With predicted masks, ADD(S)-0.1d is 58.5 / 86.2 / 83.2 / 55.6 on the four benchmarks.

Critical strengths
  • Separating localization from geometric solving is clear. In Table 3, mask IoU and pose accuracy have Spearman ρ=0.934\rho=0.934. Ground-truth masks raise YCB-Video / LineMOD from 58.5 / 55.6 to 86.0 / 88.1, so localization is the ceiling.
  • Under the default masks, GVS adds a stable 2.8–4.3 points; with worse masks it adds 10–17. With correspondences held fixed, it beats budget-matched RANSAC and TEASER++ on this correspondence set.
  • The setting needs no CAD, no task fine-tuning, and no multi-view reference set.
Critical weaknesses
  • The claimed lead of at least 28.6 points is not explained by GVS, which adds only about 3–4 points under the default masks. Table 2 mixes protocols: PoseDiffusion and RelPose++ use a sparse-view setup with ground-truth crops, Horyon is quoted from Corsetti et al. (2026), and GigaPose also receives detections and CAD.
  • If Oryon and Horyon are the direct baselines, they need to be rerun on the same anchor–query pairs. Quoted numbers cannot support a claim of p<10−75p<10^{-75} .
  • The solver is trimmed least squares. The anchor pose, depth, and intrinsics are known. Multi-instance scenes, missing depth, and large viewpoint changes are not measured.
Questions for authors
  • Please list only the methods rerun on the same 2,000 pairs, and state exactly which inputs Oryon, LatentFusion, and GigaPose received.
  • Was γ=0.7γ=0.7 tuned on these four datasets? Please plot accuracy against γγ and inlier ratio.
  • How do viewpoint gap, depth noise, and intrinsic error each change ADD(S)?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 4B8a (Part 1/2)

We thank Reviewer 4B8a for the constructive review and for highlighting our clean decoupling, strong empirical ceiling analysis (ρ=0.934\rho=0.934), and GVS's advantage over budget-matched RANSAC and TEASER++. Below we address each weakness and question point by point.

1. Exact 2,000-Pair Protocol, Inputs, and Table 2 Clarifications (W1, W2, Q1)

[W1, W2 & Q1]: Table 2 mixes protocols... If Oryon and Horyon are the direct baselines, they need to be rerun on the same anchor–query pairs... Please list only the methods rerun on the same 2,000 pairs, and state exactly which inputs Oryon, LatentFusion, and GigaPose received.

  • Methods re-run on the exact same 2,000 anchor–query pairs (seed 42): Ground2Pose (Ours), Oryon, LatentFusion, OnePoseviaGen, GigaPose, PoseDiffusion, RelPose++, and One2Any (newly added in Appendix K.1) were all executed and scored on the exact same 2,000-pair manifests per dataset using the identical evaluator. Specifically for Oryon, we verified 100% exact pair-by-pair alignment (all 2,000 instance_id entries match our manifest in content and order across all four datasets with zero missing pairs). Only Horyon (‡) is quoted from Corsetti et al. (2026) under the same dataset splits and sampling protocol because its source code and weights are not publicly available.
  • Exact inputs received by each method in Table 2:
    • Oryon: Anchor RGB-D frame (with anchor GT pose), query RGB-D frame, camera intrinsics, and either our exact Grounded-SAM mask (Prior = Ours) or GT mask (Prior = Oracle).
    • LatentFusion: Anchor/reference RGB-D view(s), query RGB-D frame, intrinsics, and identical Grounded-SAM (Ours) or GT (Oracle) masks to bound its latent 3D volume.
    • GigaPose (§): RGB images, privileged 3D CAD models (rendered into templates), per-query 2D detections/masks from Grounded-SAM (Ours) or GT (Oracle), and depth alignment for metric 6D evaluation.
    • PoseDiffusion / RelPose++ (†): Evaluated under their native sparse-view RGB protocol receiving privileged GT bounding-box crops and GT translation scale.
  • Manuscript Revision (Table 2): We added marker § to GigaPose in Table 2 and expanded the caption to explicitly define all input regimes (†, ‡, §) and clarify that the Δ row compares against the best competing practical baseline (noting where Horyon is quoted).

2. Attribution of the +28.6 to +40.8 Point Lead vs. GVS Gain (W1)

[W1]: The claimed lead of at least 28.6 points is not explained by GVS, which adds only about 3–4 points under the default masks.

  • Why off-the-shelf masks alone do not explain the lead: In Table 2, all baselines under Prior = Ours already receive our exact same Grounded-SAM masks, and under Prior = Oracle receive 100% GT masks—yet they still trail Ground2Pose by 34.7–63.2 points (and newly tested One2Any trails by +8.78 points under predicted masks and +16.99 points under GT masks). Thus, the +28.6 to +40.8 point lead stems from our decoupled match-then-verify architecture (which prevents 2D segmentation errors from corrupting 3D feature fusion, validated by 8 negative reweighting probes in Appendix C.2 & H.6).
  • Why GVS is crucial beyond the +3.5-pt mean: Under default masks, GVS delivers +17.5 to +25.8 points on difficult cluttered/symmetric objects (Tables 6–9, overall p<10−6p < 10^{-6}). Under weaker masks (Talk2DINO) or dense warp matchers (RoMa on NOCS, Appendix K.3), GVS gains surge to +10.40 to +17.80 points, acting as an essential safety shield when upstream correspondences degrade.
AuthorsOpenReview ↗

Response to Reviewer 4B8a (Part 2/2)

3. Was γ\gamma Tuned per Dataset? Accuracy vs. γ\gamma and Inlier Ratio (Q2)

[Q2]: Was γ\gamma tuned on these four datasets? Please plot accuracy against γ\gamma and inlier ratio.

  • No per-dataset tuning: γ=0.7\gamma = 0.7 (with Kiter=1,Ktop=allK_{\text{iter}}=1, K_{\text{top}}=\text{all}) was selected once on a 96-pair diagnostic probe and applied unchanged across all four datasets, all five segmenters, and all matchers.
  • Accuracy across γ∈[0.5,0.9]\gamma \in [0.5, 0.9] (Appendix H.1–H.3, Table 12): ADD(S)-0.1d is flat across a wide plateau, varying within 1.04 points across γ∈[0.5,0.9]\gamma \in [0.5, 0.9] on the YCBV sweep, and within ≤ 1.8 points (mean 0.8 points) on full 2,000-pair runs across NOCS, TOYL, and LM:
Trimming Keep Ratio γ\gamma0.50.60.7 (Default)0.80.91.0 (Untrimmed Kabsch)
YCBV Probe ADD(S)-0.1d (%)60.4260.4259.3860.4259.3854.25 (Full)
  • Accuracy vs. Effective Inlier Ratio (Table 3 & Table 14): Varying the upstream segmenter directly controls the correspondence inlier ratio entering the solver (from ~30% effective inliers under Talk2DINO up to ~90% under Oracle GT masks). Across all regimes, GVS consistently beats Kabsch and RANSAC-256, with its gain scaling inversely with inlier ratio:
Upstream Regime (Inlier Ratio Proxy)Mean Mask IoUPlain KabschRANSAC-256GVS (Ours)GVS Gain (Δ)
Talk2DINO (Low inlier ratio)44.9%39.2650.2051.86+12.60
Grounded-SAM (Moderate-high inlier ratio)70.5%67.3468.5570.88+3.54
Oracle GT Mask (High inlier ratio)100.0%84.3385.6087.68+3.35 (21.7% error reduction)

4. Viewpoint Gap, Depth Noise, and Intrinsic Error Analysis (W3, Q3)

[W3 & Q3]: ...missing depth, and large viewpoint changes are not measured. How do viewpoint gap, depth noise, and intrinsic error each change ADD(S)?

We added complete 8,000-pair viewpoint stratifications and sensor-noise sweeps in Appendix K.2 (Table 16) & Appendix K.4:

  • 1) Viewpoint Gap Stratification — 75.3% of our evaluated pairs already exceed 60°: Stratifying all 8,000 pairs by ground-truth relative camera rotation θ=arccos⁡((tr(Ra⊤Rq)−1)/2)\theta = \arccos((\mathrm{tr}(\mathbf{R}_a^\top \mathbf{R}_q)-1)/2) shows that 6,022 of 8,000 pairs (75.3%) have large viewpoint changes (> 60°), with median rotation spans of 66.5° (YCBV), 117.0° (NOCS, 100% cross-camera > 60°), 112.3° (TOYL), and 96.0° (LM):
Viewpoint Gap θ\thetaYCBV (Pairs / ADD)NOCS (Pairs / ADD)TOYL (Pairs / ADD)LM (Pairs / ADD)
Small (< 30°)557 / 71.63%0 / —286 / 95.45%145 / 66.90%
Moderate (30°–60°)391 / 57.80%0 / —233 / 90.99%366 / 66.12%
Large (> 60°)1,052 / 51.81%2,000 / 86.20%1,481 / 79.61%1,489 / 51.91%
All 2,000 Pairs2,000 / 58.50%2,000 / 86.20%2,000 / 83.20%2,000 / 55.60%

Even in the > 60° stratum, Ground2Pose reaches 51.8%–86.2%, surpassing every baseline's overall average.

  • 2) Depth Noise, Intrinsic Error & Missing Query Depth: Injecting multiplicative Gaussian depth noise σz∈{0.5,1.0,2.0}\sigma_z \in \{0.5, 1.0, 2.0\} into both depth maps before 3D lifting shifts ADD(S)-0.1d by at most −1.02 / +0.66 / −0.75 / −0.32 at σz=0.5%\sigma_z=0.5\%, −1.36 / 0.00 / −0.19 / −0.64 at σz=1.0%\sigma_z=1.0\%, and −2.04 / −1.65 / −2.44 / +0.96 (mean −1.29 pts) even at heavy σz=2.0%\sigma_z=2.0\% sensor noise. Under standard 1% camera intrinsic calibration jitter on per-frame calibrated sensors (YCBV, LM), accuracy shifts by < 0.7 points (+0.34 YCBV, −0.64 LM). Finally, when query depth is completely absent (RGB-only query), reformulating GVS into a 2D–3D reprojection verifier (GVS-PnP, Appendix K.4) achieves 31.00% / 35.40% / 27.40% / 15.95%, outperforming RGB-D Oryon and LatentFusion on YCBV and NOCS.

We hope these new experiments and clarifications fully address your concerns.

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted detailed responses to your reviews.

Please review their feedback, engage with the authors if any points remain unclear, and update your evaluation and score if their response has addressed your concerns.

Thank you

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.