acceptodds
Under review as a conference paper at ICLR 2027

AxisPose++: Reference-Anchored Tri-Axis Transport for Single-Reference Unseen Object Pose Estimation

Abstract

Estimating the unseen object's pose from minimal visual observations is essential for scalable robotic perception. While single-reference methods have shown promise, existing approaches typically rely on novel-view template synthesis, predicted geometry and local matching, or direct relative-pose regression. These pipelines remain sensitive to occlusion and weak texture and may accumulate errors through intermediate predictions. More fundamentally, the reference is primarily used as visual evidence rather than as an explicit coordinate anchor. We present AxisPose++, a reference-anchored paradigm for unseen object pose estimation from a single RGB reference. AxisPose++ formulates the task as coordinate transport: the posed reference anchors the object coordinate frame, whose compact tri-axis structure is directly transported to the target view. A geometry-grounded back-projection module then recovers the target rotation by enforcing axis orthogonality and camera-projection consistency. By avoiding costly intermediate synthesis and correspondence construction, AxisPose++ enables efficient geometry-grounded rotation recovery. Extensive experiments on synthetic and real-world benchmarks demonstrate strong generalization while being faster than the current SOTA OrienPose. Anonymous code is available at https://anonymous-pose.github.io/axispose_pp.github.io/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object pose estimation, Single-shot pose estimation, CAD model-free

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oqUgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AxisPose++, a reference-anchored framework for estimating the 3D rotation of unseen objects from a single RGB reference. Unlike prior methods that use the reference as visual evidence, AxisPose++ treats it as a coordinate anchor and reformulates the task as transporting the reference’s tri-axis structure to the target view. The method works in three stages: it first builds a coordinate-aware condition from the reference, then uses a multi-scale injection U-Net to predict both dense tri-axis likelihood fields and global axis geometry in the target view, and finally recovers the rotation by back-projecting axis hypotheses and selecting the most geometrically consistent one. Experimental results show that AxisPose++ clearly outperforms OrienPose on the ShapeNet benchmark with strong unseen-category generalization. It also transfers well to real-world benchmarks (NAVI and Objectron) without fine-tuning, while running much faster than OrienPose.

Critical strengths

++ This paper extends AxisPose, which estimates rotation by regressing a tri-axis coordinate frame, to the problem of generalizable rotation estimation from a single RGB reference. The proposed method is well-designed and effective, and the empirical results are convincing.

++ The authors provide an anonymous repository with code, training configurations, and pretrained weights, which is a plus for reproducibility.

Critical weaknesses

-- The proposed method performs somewhat worse than existing methods on the real-world benchmarks NAVI and Objectron. It would be helpful if the authors could analyze the reasons behind this gap, for example, by including some case-level discussions. In addition, the qualitative examples in Fig. 4 mostly showcase successful cases. A few examples where the method underperforms competing baselines would make the evaluation more informative.

-- Given that the method relies on a specifically defined reference coordinate frame, if the coordinate system of the reference image is altered at test time, can the relative rotation between the query and the reference still be accurately predicted? Specifically, how does altering the reference coordinate frame affect the prediction performance on the query?

Questions for authors

See above.

6 comments
AuthorsOpenReview ↗

Response to Reviewer oqUg

We sincerely thank the reviewer for the encouraging assessment and valuable suggestions. We particularly appreciate the recognition of our reference-anchored formulation, the effectiveness of the proposed method, the experimental evidence, and our efforts toward reproducibility through the anonymous code release. To be honest, we feel GENUINELY HONORED and fortunate that the effort we have devoted to this work has been understood and appreciated. We also acknowledge the concerns regarding performance gaps on real-world benchmarks and the influence of reference-coordinate changes on prediction accuracy. These are important considerations for understanding the applicability and limitations of AxisPose++. We provide detailed analyses and clarifications addressing both concerns below:

Q1. Real-world Performance and Failure Analysis. We thank the reviewer for this constructive suggestion. We acknowledge that AxisPose++ does not consistently outperform competing methods on NAVI and Objectron, with particularly noticeable performance gaps in categories such as Cup and Chair on Objectron. These limitations may arise from object symmetry, weak texture, and occlusion, which introduce ambiguity in projected axis directions, while substantial intra-category shape variations further challenge reference-anchored coordinate prediction. Although TACB and HAAS generate and evaluate multiple rotation hypotheses, they cannot fully resolve such ambiguities when discriminative visual cues are insufficient.

To address this concern, we have extended the Failure Case Analysis with four representative failure examples in Fig. 8 of the revised manuscript. These cases illustrate incorrect hypothesis selection under orientation ambiguity, missing correct candidates under severe appearance degradation, and inaccurate axis predictions caused by weak directional cues. The accompanying discussion explicitly analyzes these failure mechanisms and highlights the remaining limitations of AxisPose++.

Q2. Robustness to Reference Coordinate Changes. We thank the reviewer for this insightful question. We would like to clarify that AxisPose++ uses the posed reference to establish an explicit object-coordinate convention, rather than assuming a particular fixed coordinate orientation. Geometrically, a consistent redefinition of the object coordinate frame does not alter the relative rotation between the reference and query.

Specifically, for a coordinate-frame rotation Q∈SO(3)Q \in SO(3), the reference and target rotations transform as Rr′=RrQ⊤R'_r = R_r Q^\top and Rt′=RtQ⊤R'_t = R_t Q^\top. Their relative rotation therefore remains unchanged:

Rt′(Rr′)⊤=RtRr⊤.R'_t(R'_r)^\top = R_t R_r^\top.

In AxisPose++, the reference pose and projected tri-axis jointly define the coordinate anchor. Hence, any coordinate-frame redefinition must consistently update both representations. However, while the underlying relative rotation is geometrically invariant, exact equivariance of the learned network to arbitrary coordinate-frame changes is not explicitly enforced. Our current evaluation assumes a consistently defined reference coordinate frame. We have clarified this distinction and the corresponding applicability assumptions in the revised manuscript.

Regarding the effect of altering the reference coordinate frame at test time, could you quantify the resulting changes in the evaluation metrics and provide representative visualizations for a few samples?

Thank you for the follow-up. We have added a controlled evaluation of test-time reference frame changes to Appendix A.8 of the revised manuscript. Table 3 reports the quantitative results, and Figure 14 shows four reference-target pairs involving different objects. We also released the evaluation script and instructions on the anonymous GitHub repository for reproducing both the metrics and the visualizations.

We use the same randomly sampled 100 ShapeNet reference-target pairs in all four conditions, keeping both RGB images, camera intrinsics, and translations fixed. For each pair, we sample a 3D rotation axis and reuse it at 0°, 15°, 30°, and 60°. For a frame change QQ, we update the reference rotation to RrQ⊤R_rQ^\top and reproject its three axes. The target annotation transforms consistently to RtQ⊤R_tQ^\top, so the ground-truth relative rotation remains RtRr⊤R_tR_r^\top. We rerun the full inference pipeline for each condition and express the predicted target rotation in the original object frame before computing angular error.

Frame changeAcc@30 (%) ↑Δ Acc@30 (pp)Median angular error (°) ↓Δ median error (°)
0°69.0—10.20—
15°71.0+2.019.44+9.24
30°43.0−26.031.62+21.42
60°14.0−55.052.96+42.76

At 15°, there is a net gain of two pairs below the 30° threshold, while the median angular error increases by 9.24°. At 30° and 60°, both metrics degrade substantially. Figure 14 in the Appendix shows that individual pairs respond differently. The upper-right example degrades from 9.17° to 51.32°, the lower-left example stays below 30° in all four conditions, and the lower-right example improves from 46.56° to 12.64° at 60°. The aggregate degradation therefore coexists with substantial variation across individual pairs.

AxisPose++ uses the posed reference to make the object coordinate convention explicit and transports its tri-axis to the target view as geometric evidence for rotation recovery. When the object frame is redefined by QQ, the desired target rotation changes from RtR_t to RtQ⊤R_tQ^\top, while the RGB image pair and the physical reference-to-target rotation RtRr⊤R_tR_r^\top remain unchanged. Because the target appearance alone cannot determine the new axis convention, the estimator must recover it from the reference pose and projected tri-axis and carry it into the target prediction. Under a consistent object-frame convention, associations between appearance and axis orientation learned during training can support this transfer. A test-time redefinition changes those associations while leaving the visual evidence unchanged, requiring the predicted rotation to transform with the supplied reference frame. AxisPose++ does not explicitly enforce this equivariance, and the degradation observed at larger frame changes in Appendix A.8 shows that the learned transport can deviate from the expected transformation. The main benchmarks and component ablations still support the core contribution: using an explicit reference tri-axis to anchor rotation estimation for unseen objects under a consistent object-frame convention. The new experiment clarifies the scope of that claim by identifying sensitivity to arbitrary test-time redefinitions as a limitation of the current estimator.

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

AxisPose++ studies rotation estimation for a previously unseen object from one posed RGB reference and a target RGB image. It treats the reference pose and projected canonical axes as a coordinate anchor, predicts dense target-axis and origin likelihoods together with global 2D directions, and recovers rotation through C2-based multi-hypothesis back-projection and HAAS ranking. On the NOPE ShapeNet split, evaluated only on reference-target pairs within 45 degrees, the paper reports 68.2% Acc@30 and 15.4 degrees median error. It also reports direct transfer to NAVI and Objectron, component ablations, candidate-ranking analysis, target-image corruption tests, qualitative results, and an accuracy-efficiency comparison.

Critical strengths

The paper is clearly presented and includes fairly extensive diagnostic experiments.

Critical weaknesses
  1. The task definition is not fully consistent with the paper’s presentation. The title and introduction repeatedly use broad terms such as “object pose estimation” or “3D pose estimation,” whereas the problem formulation explicitly states that the final output is only the object rotation Rt. The paper should discuss its scope of applicability under these conditions.

  2. The dataset evaluation only reports pairs where the relative rotation between the reference image and the target image does not exceed 45 degrees. Even under a lenient metric like Acc@30, the final results are underwhelming; furthermore, the evaluation lacks metrics—such as ADD-0.1d—that account for translational error. Despite this, the paper draws broad conclusions regarding strong generalization and pose estimation for unseen objects using a single reference image—a case where the evidence does not support the claims made.

  3. Line 83 of the paper outlines the shortcomings of various baseline methods. However, the connection between this discussion and the authors' proposed approach is not clearly established. Does the authors' method actually resolve all the aforementioned issues? I am not convinced that it does. The authors should better articulate the motivation for their approach, specifically in relation to the individual modules employed.

  4. The paper claims to be 43 times faster than OrienPose; is this comparison based on evaluations conducted on the same hardware? This point requires clear, effective clarification. Furthermore, regarding the network architecture, I do not consider the design particularly efficient—especially given that it involves feature extraction using a large-scale model.

Questions for authors
  1. I am very curious about the results for pairs with a relative rotation range exceeding 45 degrees. How do precision and median error change as the target moves further away from the established reference point? Is the estimation accuracy still reliable in such cases?

  2. How are translation estimation and its accuracy achieved, and what are the metrics for translation?

7 comments
AuthorsOpenReview ↗

Response to Reviewer wpgk

We sincerely thank the reviewer for the thoughtful assessment and constructive suggestions. We appreciate the positive feedback on the overall presentation and the extensive diagnostic experiments. We understand the reviewer's concerns regarding the distinction between rotation and full 6D pose estimation, the scope of generalization claims, the limited viewpoint range, the connection between motivation and module design, and the fairness of efficiency comparisons. We have carefully addressed these concerns through task-scope clarification, additional evaluations under stricter thresholds and larger viewpoint gaps, and detailed explanations of our methodological choices and experimental protocols, as presented below:

Q1. Task definition and translation. We sincerely apologize for the ambiguity caused by our terminology, as it was not our intention to imply full 6D pose estimation. We define our study as “3D pose estimation” in the Introduction section. The term “3D pose estimation” has also been used in the literature to describe object rotation estimation, from early work in 2015 (“3d pose estimation”) [1] to recent studies such as NOPE (“3d pose estimation”, “object pose estimation”) [2] and OrienPose (“3d pose estimation”, “object pose estimation”) [3]. As explicitly defined in Sec. 3.1, we focus on estimating the target rotation (R in SO(3)), given a posed reference. Nevertheless, we acknowledge that this wording may lead to an overly broad interpretation of our task. We have clarified this scope consistently in the title, abstract, and introduction to explicitly define our research goal as “object rotation estimation”.

Q2. Evaluation Scope and Metrics. We thank the reviewer for these concerns. First, the original evaluation follows the established NOPE [2] and OrienPose [3] protocols, which restrict reference-query rotation differences to within 45° for consistent comparisons. To further examine generalization beyond this range, we conducted additional experiments with viewpoint differences up to 150°, randomly sampling 50 ShapeNet pairs per angular interval. The results are summarized below:

Relative RotationAxisPose++ Acc@30 (%) ↑OrienPose Acc@30 (%) ↑AxisPose++ Median Error (°) ↓OrienPose Median Error (°) ↓
(45°, 60°]54.032.017.2443.8
(60°, 90°]52.029.128.8963.5
(90°, 120°]44.015.456.1170.8
(120°, 150°]34.08.0375.8888.4

Notably, AxisPose++ maintains 54.0% and 52.0% Acc@30 for viewpoint gaps of 45°–60° and 60°–90°, respectively. Although performance gradually degrades under more extreme viewpoint changes, these results provide additional evidence of the method's applicability beyond the original evaluation range. All experiments use the original pretrained weights without retraining or modification. We have also released the evaluation code and sampled test pairs in our anonymous GitHub repository to facilitate reproducibility. We have incorporated these additional results in Fig.6 of the revised manuscript.

Second, we agree that Acc@30 alone is insufficient to characterize precise rotation estimation. We therefore additionally evaluated stricter thresholds on the entire ShapeNet test set:

MethodACC@5ACC@10ACC@15ACC@20ACC@25ACC@30
OrienPose0.418.532.544.351.859.1
AxisPose++ (Ours)27.945.254.860.564.868.2

As shown above, AxisPose++ consistently outperforms the OrienPose across all thresholds, with particularly substantial improvements of 27.5, 26.7, and 22.3 percentage points at ACC@5, ACC@10, and ACC@15, respectively. Notably, the performance advantage becomes more pronounced under stricter thresholds, demonstrating that our improvements are not an artifact of the relatively permissive ACC@30 metric. We have incorporated these additional threshold-wise results in Table 2 of the revised manuscript.

Finally, as clarified in Q1 and explicitly defined in Sec. 3.1, AxisPose++ estimates only the rotational component (Rt∈SO(3)R_t \in SO(3)), rather than a 6D pose with translation. Therefore, translation-dependent metrics such as ADD-0.1d are not applicable to our current task. We acknowledge that our conclusions should be restricted to rotation estimation, and we have clarified the evaluation scope and limitations in the revised manuscript.

Q3. Motivation and Module Design. We thank the reviewer for this constructive comment. We would like to clarify that our method does not claim to completely resolve all limitations of existing approaches. Rather, AxisPose++ introduces a reference-anchored formulation to alleviate its key dependencies through three complementary stages.

Specifically, (1) RCA addresses the reliance on appearance-based reference representations by explicitly incorporating the known reference pose and projected tri-axis to establish a coordinate anchor. (2) T³, equipped with MCCI and TLF, directly predicts target-view tri-axis evidence conditioned on this anchor, bypassing intermediate novel-view synthesis, depth estimation, and explicit cross-view feature matching. (3) GGPR, consisting of TACB and HAAS, addresses ambiguous axis predictions by constructing multiple geometrically constrained rotation hypotheses and selecting them using projection consistency and likelihood support.

Together, these stages reduce dependence on potentially unreliable intermediate representations and avoid expensive per-query synthesis and matching, while achieving a 43× speedup over OrienPose. Nevertheless, challenges such as intrinsic symmetry and severe occlusion remain, as discussed in our failure analysis. We have revised the Experiment section to more explicitly connect each limitation with the corresponding design motivation, without implying that all challenges are completely resolved.

Q4. Inference Efficiency. We thank the reviewer for this concern. We would like to clarify that both AxisPose++ and OrienPose were evaluated end-to-end on the same NVIDIA RTX 4090 GPU under consistent experimental conditions. The measurements include the complete inference pipelines, covering feature extraction, tri-axis prediction, and geometric hypothesis selection for AxisPose++, as well as novel-view synthesis and template comparison for OrienPose. Under this setting, AxisPose++ achieves 2.75 FPS versus 0.063 FPS for OrienPose, corresponding to a 43× speedup.

Regarding model complexity, AxisPose++ employs a frozen small DINOv2 ViT-S/14, rather than a large-scale ViT backbone. More importantly, our efficiency advantage stems from the overall inference formulation: AxisPose++ directly predicts compact tri-axis representations and recovers rotations algebraically, avoiding the repeated novel-view synthesis and template comparisons required by generate-and-compare methods. Thus, despite the computational cost of feature extraction, the complete pipeline achieves substantially lower inference latency. We have clarified the timing protocol and the source of the efficiency gain in the revised manuscript.

Q5. Generalization to Large Viewpoint Gaps. We thank the reviewer for this valuable question. As answered in Q2, the original evaluation follows the established NOPE [2] and OrienPose [3] protocols, which restrict relative rotations to within 45°for consistent comparisons. To further investigate performance under larger viewpoint changes, we conducted additional experiments extending the relative rotation range to 150°, and we shared the evaluation code and sampled test pairs in our anonymous GitHub repository to facilitate reproducibility.

Q6. Translation Estimation. We thank the reviewer for this question. As clarified in Q1 and explicitly defined in Sec. 3.1, AxisPose++ focuses exclusively on 3D rotation estimation, rather than full 6D pose recovery. Although TACB introduces an auxiliary translation th=λ0n(cH)\mathbf{t}^h = \lambda_0 \mathbf{n}(\mathbf{c}_H) during geometric back-projection, λ0\lambda_0 is an arbitrary reconstruction scale. Therefore, this auxiliary variable does not represent metric object translation and is not retained as the final output. Moreover, our training objectives do not include translation supervision. Accordingly, we evaluate rotation accuracy using Acc and median angular error, rather than translation-dependent metrics such as ADD-0.1d. We have clarified this distinction in the revised manuscript.

Reference:

[1] Wohlhart, Paul, and Vincent Lepetit. "Learning descriptors for object recognition and 3d pose estimation." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.

[2] Nguyen, Van Nguyen, et al. "Nope: Novel object pose estimation from a single image." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024.

[3] Liu, Yating, et al. "OrienPose: Orientation-Guided Novel View Synthesis for Single-Image Unseen Object Pose Estimation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

Thank you again for your insightful review. In the revised manuscript, we clarified that AxisPose++ estimates 3D object rotation from a single posed reference and added a ShapeNet evaluation across four relative rotation ranges beyond 45°. AxisPose++ achieves higher Acc@30 and lower median angular error than OrienPose in all four ranges, although both methods degrade as the viewpoint gap grows. We also added results at stricter angular error thresholds to show how the comparison changes when finer rotation accuracy is required. If any concerns remain, we would be happy to clarify the evaluation or run additional targeted experiments where they would help address a specific question.

Reviewer_33CNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
3(+1)
Read review
Paper summary

Changed Presentation from 1 to 2 and soundness from 2 to 3. LGTM now!


The authors propose AxisPose++, which mainly consists of reference-coordinate anchoring (RCA), target-view tri-axis transport (T^3), geometry-grounded pose recovery (GGPR). Through these modules, authors suggest the easy way to tackle single-reference rotation estimation problem, especially using hypotheses-based optimal axis value selection, called hypothesis-aware axie selection (HAAS).

The authors well-summarize/categorize the existing baseline approaches, but my main concern of this paper is two fold. i) overclaim about the generalization capability: if I understand corectly, it's not "pose estimation", but rather it's rotation estimation. And authors highlight generalization capability, but I don't see the significant generalizaiton capability. Especially, in general, if one wants to argue generalizabilty, the cross-data evaluation or zero-shot estimation should be presented, but neither was provided. Plus, in NAVI and Objectron datasets, their performance increase is not significant — if the performance itself is poor, can we say it's generalizable? ii) weak criteria of evaluation: if rotation error is within 30 deg, an estimate is considered succeeded. But from the robotics researchers' perspective, a 30-degree error cannot be considered a success. So I'm not sure, even in ShapeNet, the author could say it's better. Imagine that, there exist two rotation errors: [1, 1, 1, 1, 1, 1, 1, 90] vs. [20, 20, 20, 20, 20, 20, 20, 20] errors, is that fair that latter one is better? Might worth drawing success rate plot along the thresholds (i.e., x axis: X of Acc@X, and y axis: Acc@X)

Critical strengths
  • To enable robust estimation, using hypotheses really makes sense.
  • thorough ablation studies in Fig. 6(a) support that each module is necessary to achieve high performance.
Critical weaknesses
  1. Limited performance on real-world datasets: The proposed method shows relatively poor performance on real-world datasets, particularly NAVI and Objectron. This raises concerns about its practical applicability beyond controlled experimental settings.

2.Insufficient evidence of generalizability: Although the paper emphasizes generalizability, the experimental results on NAVI and Objectron do not convincingly support this claim. In fact, the performance degradation on these datasets suggests potential challenges in transferring the proposed approach to real-world scenarios.

  1. Lack of clarity in the methodology: Sections 3.3 and 3.4 are difficult to follow, primarily due to the lack of visual explanations. An additional figure or schematic illustration would substantially improve the clarity and accessibility of the proposed methodology.
Questions for authors

Q1. In Multi-scale Coordinate-Control Injection, why is the scale added? Intuitively, I would expect the scale to be multiplied rather than added.

Q2. You mentioned that "As illustrated in Figure 3, weak texture or object symmetry may induce multiple plausible hypotheses for one or more axes." I suspect this may explain the poor performance of AxisPose++ on the Mug and Cup categories. The authors should clarify whether HAAS can effectively handle multiple symmetry-equivalent rotations.

Q3. The paper should be self-contained, but the current manuscript does not sufficiently meet this expectation. For example, readers who are unfamiliar with the C2 formulation should be able to understand its underlying mechanism solely from this manuscript. Unfortunately, the current explanation does not provide sufficient detail.

Q4. What do the red and gray dots on the right side of Fig. 6(b) represent? The figure is difficult to follow without a legend.

Q5. The claim of being "43× faster" seems somewhat overstated. Is this speedup based on a comparison with OrienPose, one of the slowest baselines? It seems unfair to highlight the speedup relative to the slowest method. In addition, the authors should clarify how the inference times were measured. If I understand correctly, the methods involve different preprocessing steps, and some timing results may have been taken from the original papers rather than measured on the same hardware under consistent experimental conditions.

9 comments
AuthorsOpenReview ↗

Response to Reviewer 33CN

We sincerely thank the reviewer for the constructive feedback and careful evaluation of our work. We appreciate the reviewer's recognition of our systematic categorization of existing approaches, the motivation for hypothesis-based reasoning, and the comprehensive ablation studies supporting our design. We also acknowledge the concerns regarding generalization capability, evaluation thresholds, methodological clarity, and inference efficiency. We have carefully considered these comments and provide detailed clarifications, additional experimental results, and point-by-point responses to each concern below:

Q1. Task scope and generalization:First, we sincerely apologize for the ambiguity caused by our terminology, as it was not our intention to imply full 6D pose estimation. We define our study as “3D pose estimation” in the Introduction section. The term “3D pose estimation” has also been used in the literature to describe object rotation estimation, from early work in 2015 (“3d pose estimation”) [1] to recent studies such as NOPE (“3d pose estimation”, “object pose estimation”) [2] and OrienPose (“3d pose estimation”, “object pose estimation”) [3]. As explicitly defined in Sec. 3.1, we focus on estimating the target rotation (R in SO(3)), given a posed reference. Nevertheless, we acknowledge that this wording may lead to an overly broad interpretation of our task. We have therefore revised the title, abstract, introduction, and conclusion to consistently clarify the scope of our work as object rotation estimation rather than full 6D pose estimation.

Second, as for the concern of generalization capability. In the domain of object pose estimation, generalization is commonly evaluated at two levels [4]: generalization to unseen instances within training categories (category-level generalization) and the more challenging generalization to objects from entirely unseen categories (unseen-object level generalization). Our work explicitly targets the latter under a particularly constrained setting, where only a single posed RGB reference is available, without CAD models, depth observations, or multiple reference views. Specifically, we train AxisPose++ on 11 ShapeNet categories and evaluate it on 6 entirely unseen categories, achieving 68.2% Acc@30 without category-specific adaptation. Furthermore, we directly evaluate the ShapeNet-trained model on the real-world NAVI and Objectron datasets without any fine-tuning, providing additional evidence of zero-shot cross-dataset generalization, which was also acknowledged by reviewer oqUg:“AxisPose++ clearly outperforms OrienPose on the ShapeNet benchmark with strong unseen-category generalization.” We acknowledge that generalizability does not necessarily imply superior performance on every benchmark, so we have clarified this distinction in the Experiment section to avoid overstating our claims.

Q2. Evaluation thresholds. We thank the reviewer for this valuable suggestion. We agree that Acc@30 alone cannot fully characterize rotation accuracy, particularly for applications requiring precise orientation estimation. We adopt Acc@30 following the established evaluation protocols of NOPE [2] and OrienPose [3], rather than treating a 30°error as sufficient for robotic applications. To provide a more comprehensive evaluation, we additionally report Acc@X under stricter thresholds on the entire ShapeNet unseen-category test set in Table 2 of the revised manuscript:

MethodACC@5ACC@10ACC@15ACC@20ACC@25ACC@30
OrienPose0.418.532.544.351.859.1
AxisPose++ (Ours)27.945.254.860.564.868.2

As shown above, AxisPose++ consistently outperforms the OrienPose across all thresholds, with particularly substantial improvements of 27.5, 26.7, and 22.3 percentage points at ACC@5, ACC@10, and ACC@15, respectively. Notably, the performance advantage becomes more pronounced under stricter thresholds, demonstrating that our improvements are not an artifact of the relatively permissive ACC@30 metric. We have incorporated these additional threshold-wise results in Table 2.

Q3. Real-world Performance. We thank the reviewer for this concern and acknowledge that AxisPose++ does not consistently outperform all competing methods on real-world benchmarks. However, we would like to emphasize that our evaluation follows a strict zero-shot domain transfer setting: AxisPose++ is trained exclusively on synthetic ShapeNet data and directly evaluated on NAVI and Objectron without any real-world fine-tuning or adaptation. Under this setting, AxisPose++ achieves the second-highest Acc@30 on NAVI (47.1% / 33.5° median error) and a tied second-highest Acc@30 on Objectron (53.9% / 29.8°). Compared with NOPE, this represents improvements of 10.3 and 3.1 percentage points in Acc@30, respectively. These results demonstrate cross-dataset transferability, although they do not imply consistently superior real-world performance. Importantly, AxisPose++ also offers a substantial efficiency advantage, achieving 2.75 FPS compared with 0.063 FPS for OrienPose, corresponding to a 43× speedup under end-to-end evaluation on the same hardware. This highlights a favorable accuracy-efficiency trade-off for practical deployment.

Q4. Generalization Capability. We thank the reviewer for this concern. We would like to clarify that the generalizability of AxisPose++ is evaluated at two complementary levels. First, unseen-category generalization is evaluated on ShapeNet, where the model is trained on 11 categories and tested on six entirely unseen categories, achieving 68.2% Acc@30 without category-specific adaptation. Second, zero-shot cross-dataset generalization is evaluated by directly transferring the ShapeNet-trained model to real-world NAVI and Objectron without any fine-tuning. As discussed in Q3, AxisPose++ achieves competitive results on both benchmarks, including improvements of 10.3 and 3.1 percentage points in Acc@30 over NOPE, respectively. Notably, lower performance on real-world benchmarks is not unique to our method. Representative baselines such as OrienPose (59.1% on ShapeNet vs. 50.9%/53.5% on NAVI/Objectron) and NOPE (52.5% vs. 36.8%/50.8%) also exhibit performance differences across these benchmarks, reflecting the challenges of cross-dataset evaluation. Despite these challenges, AxisPose++ achieves the second-highest Acc@30 on NAVI (47.1%) and ties for second on Objectron (53.9%), outperforming NOPE by 10.3 and 3.1 percentage points, respectively.

These results provide evidence of generalization beyond the training categories and synthetic domain, although we acknowledge that real-world domain shifts remain challenging. We have clarified the generalization and discussed the limitations more explicitly in the Failure Case Analysis section of the revised manuscript.

Q5. Methodological Clarity. We thank the reviewer for this suggestion. We would like to clarify that the manuscript already provides complementary explanations of the proposed methodology. Specifically, Fig. 2 illustrates the overall architecture and information flow, while Fig. 3 visualizes tri-axis ambiguity and hypothesis selection. Furthermore, Appendices A.2–A.3 provide detailed mathematical derivations of the C2-based back-projection and HAAS selection criteria, together with the complete pseudocode in Algorithm 1. To further support reproducibility, we have also released an anonymous repository containing the complete training and inference code, environment configurations, pretrained weights, and evaluation data. We believe these materials provide a comprehensive description of the proposed method.

Q6. Additive Conditioning in MCCI. We thank the reviewer for this question. We would like to clarify that “multi-scale” refers to different spatial resolutions in the feature hierarchy, rather than numerical scaling factors. As described in Sec. 3.3, MCCI injects reference-coordinate information into the target features through residual addition:

X~ts=Xts+Zs(Crs),\tilde{X}_t^s = X_t^s + Z_s(C_r^s),

where ZsZ_s is a zero-initialized 1×11 \times 1 convolution. This design preserves the original target feature pathway at initialization and allows reference-coordinate information to be progressively incorporated during training. While multiplicative conditioning is also feasible, it directly modulates target feature activations, whereas our additive formulation introduces complementary reference information without explicitly suppressing the target features.

Q7. Symmetry Ambiguity. We thank the reviewer for this insightful question. HAAS is designed to handle ambiguous axis predictions by constructing multiple geometrically valid rotation hypotheses and selecting the one best supported by directional reprojection consistency and predicted tri-axis likelihoods. When discriminative visual cues are available, this mechanism can distinguish among competing orientations. However, as discussed in the Failure Case Analysis section, we acknowledge that intrinsically symmetry-equivalent rotations cannot be uniquely distinguished when their visual observations are indistinguishable. This limitation may partly explain the lower performance on Mug and Cup, particularly when symmetry-breaking cues, such as handles or distinctive textures, are occluded or weak. Therefore, HAAS alleviates ambiguity arising from uncertain axis predictions but does not fully resolve inherent rotational symmetry. We have clarified this distinction and discussed the corresponding failure cases in the Failure Case Analysis section of the revised manuscript.

Q8. Self-contained C2 Formulation. We thank the reviewer for this comment. We would like to clarify that Appendix A.2 (Eqs. 8–17) already provides a complete, self-contained derivation of the C2-based back-projection used in AxisPose++. Specifically, we first lift the predicted 2D tri-axis configuration into 3D camera rays (Eqs. 8–11), then enforce pairwise orthogonality among the recovered axes (Eq. 12). These constraints are algebraically reduced to a quadratic equation with explicitly derived coefficients (Eqs. 13–15). Finally, the recovered 3D axes are normalized and projected onto SO(3) to obtain valid rotation hypotheses (Eqs. 16–17). Thus, the complete recovery procedure can be understood and reproduced without consulting the original C2. Also, we have released the complete training and evaluation code, pretrained weights, and a test set, including a detailed implementation of the C2 formulation.

Q9. Clarification of Fig. 6(b). We thank the reviewer for pointing this out. Each point represents one test pair, with the horizontal and vertical coordinates indicating the rotation errors selected by likelihood-only ranking and HAAS, respectively. Red points indicate cases where HAAS achieves a lower rotation error, while gray points indicate cases without improvement. The dashed diagonal denotes equal errors between the two methods; thus, points below the diagonal indicate improvements from HAAS. We have added a clear legend and corresponding explanations to Fig. 6(b) in the revised manuscript.

Q10. Inference Efficiency. We thank the reviewer for this concern. We would like to clarify that OrienPose [3] (CVPR 2026) was selected not because of its relatively slow inference speed, but because it is a recent and highly competitive method under the same single-reference RGB-only setting, achieving the strongest performance among existing baselines on our main ShapeNet benchmark. Therefore, comparing against OrienPose provides a meaningful assessment of the accuracy-efficiency trade-off. Importantly, both methods were evaluated end-to-end on the same NVIDIA RTX 4090 GPU under consistent experimental conditions. The measurements cover the complete inference pipelines, including novel-view synthesis and template comparison for OrienPose, and feature extraction, tri-axis prediction, and geometric hypothesis selection for AxisPose++. Under this setting, AxisPose++ achieves 2.75 FPS versus 0.063 FPS for OrienPose, corresponding to a 43× speedup while improving Acc@30 from 59.1% to 68.2%. Furthermore, Fig. 1(f) compares the accuracy-efficiency trade-off across multiple baselines, rather than relying solely on this pairwise comparison. We have clarified the timing protocol in the revised manuscript.

Reference:

[1] Wohlhart, Paul, and Vincent Lepetit. "Learning descriptors for object recognition and 3d pose estimation." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.

[2] Nguyen, Van Nguyen, et al. "Nope: Novel object pose estimation from a single image." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024.

[3] Liu, Yating, et al. "OrienPose: Orientation-Guided Novel View Synthesis for Single-Image Unseen Object Pose Estimation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[4] Liu, Jian, et al. "Deep Learning-Based Object Pose Estimation: A Comprehensive Survey: J. Liu et al." International Journal of Computer Vision 134.2 (2026): 81.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.