acceptodds
Withdrawn from ICLR 2027

Mind the Orientation: Overcoming UprightBias in Feed-Forward Geometric Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We uncover a largely overlooked vulnerability in feed forward reconstruction models: sensitivity to image orientation. Across state-of-the-art architectures rotational perturbations of input images cause severe degradation in both camera pose estimation and point-map quality. We term this failure mode UprightBias and trace its origin to a near-universal upright orientation of training corpora, which causes models to implicitly assume canonically upright inputs at inference time. We find that patch-level embedding variance from a DINOv2 encoder provides a reliable zero-shot signal for identifying the canonical upright orientation. Leveraging this insight, we introduce ROVER (ROtation estimation via Variance-based Embedding Rectification), a training-free framework that estimates and rectifies image orientation at inference, leaving all downstream model weights unchanged. For evaluation, we introduce a Controlled Rotation Synthesis protocol that generates rotated views from upright images. We also curateRotMV (Rotational Multi-View), which, to the best of our knowledge, is the first benchmark of paired upright and real-world rotated multi-view sequences captured from fixed viewpoints. Experiments across the datasets demonstrate that ROVER consistently shows improvements, reducing pose estimation errors upto 80% and point-map estimation errors by over 58% on average.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Feed Forward Models, 3D Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Aw8DOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates an overlooked failure mode in feed-forward 3D reconstruction models, termed UprightBias, where arbitrary in-plane rotations of input images can substantially degrade camera pose estimation and geometric reconstruction. The authors demonstrate this phenomenon across several recent models, including VGGT, VGGT-Ω, π³, MapAnything, and Depth Anything 3.

To mitigate this issue, the paper proposes ROVER, a training-free, plug-and-play approach that estimates the canonical upright orientation of each input image by maximizing the spatial variance of DINOv2 patch embeddings. The inputs are then rotated back to their estimated upright orientations before being processed by the original reconstruction models. The authors also introduce a controlled rotation synthesis pipeline and a real-world rotational multi-view benchmark (RotMV). Experiments show substantial improvements in reconstruction quality and camera pose estimation under rotated inputs.

Critical strengths
  1. An interesting and practically relevant observation. The paper identifies a meaningful robustness issue in modern feed-forward geometric models that is not adequately captured by conventional reconstruction benchmarks. The results highlight the importance of evaluating geometric models beyond standard, predominantly upright image distributions.

  2. Extensive empirical evaluation. The authors investigate UprightBias across multiple model architectures and datasets, including both synthetically rotated images and real-world captures. The consistency of the observed degradation across different models provides convincing evidence that this is a widespread issue.

  3. A simple and effective solution. ROVER is lightweight, training-free, and compatible with existing reconstruction models without architectural modifications. Despite its simplicity, it demonstrates substantial improvements in both camera pose estimation and reconstruction quality.

Critical weaknesses
  1. The treatment of rotation-induced artifacts during inference is unclear. The authors carefully avoid padding and cropping artifacts when constructing their synthetic evaluation data by rendering rotated views using 3DGS. However, the paper does not sufficiently explain how these artifacts should be handled during inference.
  2. Limited robustness of the proposed orientation estimation. ROVER relies on the assumption that the spatial variance of DINOv2 patch embeddings is maximized at the canonical upright orientation, which may not hold consistently across different scene configurations. In fact, the authors report a failure case on the ETH3D office scene, where unusual viewpoints and scene symmetry lead to incorrect orientation estimates and substantially degraded reconstruction. Moreover, the supplementary results show non-negligible performance drops on some already-upright inputs, despite the claims of negligible degradation in the main text. A more systematic evaluation of orientation estimation failures, along with a mechanism to avoid unnecessary corrections when the estimated orientation is unreliable, would improve the practical robustness of the method.
Questions for authors

none

7 comments
AuthorsOpenReview ↗

Robustness discussion added, Non-negligible drop in performance on minor case discussed

Rotation-induced artifacts during inference

At inference, ROVER applies the estimated correction with a single in-plane rotation of the input image using simple rotation transformation with bilinear interpolation (L 230). This is a design choice: it keeps inference feed-forward, adds no per-scene processing, and costs about 0.43sec for 16 images (Table H). Controlled Rotation Synthesis (Section 3.1) requires fitting a 3DGS model per scene, so it is an evaluation tool and cannot be part of a feed-forward pipeline during inference as it requires test time 3DGS fitting and optional Difix3d+. The rotation-induced padding is already included in every ROVER result. The rotated inputs in our benchmarks are rendered without padding (Section 3.1). Despite this, ROVER improves over the rotated-input baseline across five models and seven datasets, including real captures in RotMV (Tables 3, 4, A, B, C, D, E), with AUC$_{30}$ gains of about 20 points at 100% rotation (Figures 4, D, E, F; Table 3, 4, B, C, D, E). We have added this implementation detail and its rationale to the revised manuscript. Thus during inference we rotate the image and pad it accordingly

Robustness of orientation estimation

ROVER achieves the lowest orientation error among all evaluated alternatives, with a mean error of 10.3° on NRGB-D, compared to 39.1° for RotNet and 56.3° for GeoCalib (Table 5). The same ordering is observed on 7-Scenes and Tanks & Temples (Table G). Furthermore, the reported failure is rare, occurring in only one of the 13 ETH3D scenes (Supp. Section I), and we observe no comparable degradation on any of the other evaluated datasets, including NRGB-D, Tanks & Temples, Mip-NeRF, 7-Scenes, and DTU. We report this failure case to maintain scientific transparency and provide an honest account of the method's limitations. However, this isolated occurrence should be interpreted in the context of ROVER's consistent performance across the broader evaluation, rather than as a critical weakness of the proposed approach.

Non-negligible drop in performance drops on minor case

We clarify that by negligible performance degradation, we mean that the observed differences between the baseline and ROVER fall within the measurement variance of the reconstruction metrics. This is consistent with our main-paper results, where N.C. is preserved across all models and changes in AUC₃₀ remain within 0.02 points. We further extend this evaluation across multiple datasets in Supp. Table F. Averaged across these datasets, four of the five models (VGGT-Ω, VGGT, DA3, and π³) exhibit changes in N.C. of at most 1.3% (e.g., VGGT-Ω: 0.73 →\rightarrow 0.73; DA3: 0.72 →\rightarrow 0.73) and changes in AUC₃₀ of at most 1.2% points (VGGT-Ω: 96.7 →\rightarrow 95.8; VGGT: 91.6 →\rightarrow 91.1; DA3: 97.4 →\rightarrow 96.5; π³: 90.7 →\rightarrow 90.1). MapAnything also exhibits negligible variation in N.C. (0.65 →\rightarrow 0.66), although its AUC₃₀ decreases from 81.64 to 78.58. We report this exception explicitly. Since MapAnything predicts metric-scale 3D reconstructions, this degradation may be related to its metric-scale reconstruction formulation and appears to be model-specific and investigation for this is beyond the scope of this manuscript.

Summary

We clarified that ROVER's inference-time correction is a single, low-cost in-plane rotation, and that its gains hold despite the padding this introduces, which addresses the concern about rotation-induced artifacts. We also added comparisons showing ROVER has the lowest orientation error among the alternatives which addresses the concerns about estimation robustness and performance on upright inputs.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer Aw8D

With about a week remaining in the discussion period, we wanted to kindly invite you to take another look at our rebuttal and the updated results we have added based on your valuable feedback.

We have addressed all of your comments in detail, added sections and tables, clarified rotation-induced artifacts during inference, added discussion on robustness of orientation estimation and addressed the supplementary table results on all upright images. We would be very grateful if you could share any further thoughts, or let us know if there is anything else we can clarify.

Thank you again for your time and for helping us improve the paper.

Best regards,

Authors

Thanks for the rebuttal. I appreciate the authors' efforts to address my concerns. The clarification regarding inference-time padding is helpful. However, as I commented in Reviewer HLYf's thread, the upright bias is not the only "bias" in real-world image distributions, and you cannot fix all of them into a "canonical" form. Rotating and cropping further introduce information loss and black regions that may cause serious problems for downstream models that contain, e.g., "black region bias" (P.S., this is true for video generators that you better not have any black pixels in the inputs). Furthermore, an error of 10.3° is not good enough (although the proposed model outperforms the baselines) to really solve the problem. From the reviewer's point of view, the harm it introduces might be larger than the benefits it brings to the system.

Based on these reasons, I would like to keep my scores unchanged.

Reviewer_HLYfOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper quantitatively identifies the feed-forward networks' bias: they are trained on mostly upright data, so they perform well on upright data and suffer from degraded performance on rotated data.

To deal with the bias, the authors propose to find the canonical orientation (aka gravity) of a rotated image, and unrotate the image in the pixel space. The orientation is identified by spatial variance of patch embeddings with the DINOv2 extractor. It is thus computed by brute force searching angles degree-by-degree, and accelerated by the coarse-to-fine searching.

To systematically study the bias, authors create a synthetic dataset by synthesizing rotated novel views through 3DGS reconstruction, enhanced by Difix3D+. They also introduced a new dataset, RotMV, to check the applications to real images.

Critical strengths
  • The paper did due diligence to systematically study the upright bias and identified a real data distribution pattern in general training data.
  • It also introduces a new multi-view dataset with rotated images.
Critical weaknesses

At the high level, I think the problem setup needs a better motivation.

  • It is reasonable that the training data are by construction upright, because it is natural. Human, and most mammals' visual systems are by construction upright. The same applies to the pretrained ViT backbones.
  • In real life, humans see upright, users hold phones upright, photographers create pictures upright, vehicle dash cameras are upright. When I use VGGT, I would not deliberately rotate my personal collection of images by 195 degrees and feed it into the network. I am just fine with a biased network dealing with naturally biased daily data.
  • In my opinion, applied science should not be asking a question for the sake of asking a question. I don't see any valid real-world applications in this paper where we HAVE to deal with rotated images. It should be the author's responsibility to identify at least some real-world scenarios where the proposed approach can be applied. The only persistent non-upright data I can think of is drone data, but I suspect VGGT-Omega works fine on this modality, since the drone scene is one of its demo cases.

The methodology is also quite flawed.

  • To probe the upright bias, the authors created synthetic datasets from existing datasets through 3DGS rendering and diffusion-based enhancement, which intrinsically creates further bias: imperfect 3DGS rendering and generated image restoration are unseen by the feed-forward models as well. In the specific setup, it's hard to tell whether the degraded performance is really from orientation, or indeed, synthesis-to-real gap.
  • In this regard, the actual valuable contribution and plausible analysis are upon the RotMV dataset. But unfortunately, the poses are estimated by VGGT: it is slightly circular argument, basically you acknowledge the estimation from upright (0 degree) as GT, then conduct analysis to show that rotated estimates diverge from it.
  • To clarify, I'm not saying the conclusions are wrong from those experiments, they are possibly true. My point is that they are not rigorous and not convincing.

The ROVER approach is a bit raw through brute force (although accelerated) angle search. There must be more concise approaches to do so (e.g. a lightweight MLP).

Questions for authors
  • I would like the authors to at least find one valid use case / benchmark in the real world (not limited to reconstruction, but any practical use case), and build concrete evidence that current approaches would fail, and the only rescue is ROVER. Otherwise I will define it as a good case study and recommend a submission to a workshop.
9 comments
AuthorsOpenReview ↗

Syn-to-Real Gap experiment added, COLMAP estimation for RotMV included, Trained MLP Baseline reported

Influence of the rendering quality on the final results (Syn-to-Real gap)

We account for the influence of rendering quality through a (1) multi-stage quality-control pipeline and (2) an additional quantitative validation experiment introduced in response to the reviewer’s suggestion. We perform extensive quality control (Supp. Section H) to minimize rendering artifacts. First, we fit a state-of-the-art 3DGS model to the scene, enabling novel-view synthesis. We then optionally apply the DifiX3D+ (Figures J–K), followed by manual inspection and filtering of the generated views (Figure L). This process ensures that the evaluated synthetic images are of high visual quality and minimizes potential confounding effects from rendering artifacts. To isolate the effect of rendering quality, we conduct an additional experiment in which we synthesize upright (0) images using Controlled Rotation Synthesis (sec 3.1), retaining the original camera poses from which the images were captured (Supp. Section N; Table R). We then evaluate downstream performance against the original images to determine whether the rendering process itself affects reconstruction quality, independently of camera rotation. The results show that the rendering process introduces negligible changes in downstream performance, with differences remaining within the observed variance across runs. This suggests that the rendering process itself does not materially affect reconstruction quality.

Using VGGT-estimated upright poses for RotMV

Following the reviewer's suggestion, we re-estimate the camera pose using an independent SfM pipeline (COLMAP) instead of VGGT. The corresponding results are reported in Table S (Supp. Section O). We observe consistent downstream performance, with differences remaining within the observed variance, on the RotMV dataset across both pose estimation pipelines. Indicating that our findings are not sensitive to the choice of camera pose estimator. Also to clarify a bit in our original protocol, the upright image at each viewpoint was known in the collected data. We therefore used VGGT exclusively on these known upright images to estimate the reference camera poses. This was justified because using a learned pose estimator as an external pose oracle is an established practice. For instance, [1] explicitly uses VGGT as their pose oracle in place of COLMAP, citing its robustness and ease of use. This choice is also relevant in sparse-view settings, where conventional SfM can be unreliable due to limited feature correspondences [2]. Once the reference poses were obtained, we derived the poses for the remaining roll angles analytically by applying the corresponding known in-plane rotations to the reference camera orientations, while keeping the camera centers and viewing directions fixed. Thus, the rotated camera poses were determined geometrically rather than estimated by VGGT from the rotated images.

References

[1] Mitchel, T. W., Ryu, H., & Sitzmann, V. (2025). True self-supervised novel view synthesis is transferable. ICLR 2026

[2] Fan, Z., Wen, K., Cong, W., Wang, K., Zhang, J., Ding, X., Xu, D., Ivanovic, B., Pavone, M., Pavlakos, G., Wang, Z., & Wang, Y. (2024). InstantSplat: Sparse-view Gaussian splatting in seconds.

Trained MLP Baseline

We have investigated an alternative approach based on direct rotation-angle regression using a lightweight MLP prediction head (Supp. Section M). In this experiment, the MLP is trained to predict the angle of the rotation angle from the extracted image features. However, this approach performs less effectively than ROVER's variance-based estimation. We hypothesize that the features of a single image do not encode its absolute orientation in a form that a small head can decode: orientation must be inferred from scene content, so regression is sensitive to scene type and to the training distribution. ROVER compares the spatial variance of the same image's embeddings across candidate rotations and selects the maximum, which is a relative comparison within each image and needs no learned mapping from features to angle.

AuthorsOpenReview ↗

Real World Use Case

Real World Use Case for ROVER

Orientation robustness matters in settings where camera roll is not controlled, such as egocentric capture (AR/VR glasses, body cameras) and robotics, where head or arm motion produces large, uncontrolled roll that can degrade reconstruction models.

To test a real setting, we evaluated VGGT- with and without ROVER on EPIC-Kitchens (dynamic, egocentric video) [1], using the COLMAP camera poses provided with the dataset as reference. We observed that ROVER improves camera pose estimation compared to VGGT-Omega. This concludes that ROVER has a valid use case in the real world where current state of the art approaches provide suboptimal reconstruction and ROVER rescues it.

Orientation estimation is a well-studied problem in single-image settings. It is treated as "gravity-direction estimation" [2, 3, 4], and is evaluated on real AR and indoor datasets such as LaMAR and Stanford2D3D. In monocular depth estimation it appears as the "horizontal prior"[5], where DIODE and NYUv2 have been used to measure how depth models degrade under camera roll. We study the multi-view counterpart in feed-forward reconstruction, which we call UprightBias; to our knowledge, it has not been characterized for these models. Quantifying it on real data is difficult because no existing multi-view benchmark provides paired upright and rotated captures, which motivates our RotMV benchmark.

Epic-Kitchen

MetricWithout ROVERWith ROVER
Tacc@30 (avg. of 2 windows)15.5%18.3% (+2.8 pts)
Racc@10 (1 window)59.3%60.3% (+1.0 pt)

References

[1]: Damen, Dima, et al. "The epic-kitchens dataset: Collection, challenges and baselines." IEEE Transactions on Pattern Analysis and Machine Intelligence 43.11 (2020): 4125-4141.

[2]: Veicht, Alexander, et al. "Geocalib: Learning single-image calibration with geometric optimization." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[3] Tirado-Garín, Javier, and Javier Civera. "AnyCalib: On-manifold learning for model-agnostic single-view camera calibration." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

[4]: Lochman, Yaroslava, et al. "Babelcalib: A universal approach to calibrating central cameras." 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2021.

[5]: Tang, Kaihua, et al. "Breaking the Horizontal Prior: From Long-Tailed Orientation Bias to Roll-Robust Monocular Depth Estimation." arXiv preprint arXiv:2608.00678 (2026).

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer HLYf

With about a week remaining in the discussion period, we wanted to kindly invite you to take another look at our rebuttal and the updated results we have added based on your valuable feedback.

We have addressed all of your comments in detail, updated several figures and tables, added real world use case for ROVER, included trained MLP basline to estimate orientation, provided RotMV results with COLMAP estimated geometry and experiments to rule out syn-to-real gap as possibilities for results. We would be very grateful if you could share any further thoughts, or let us know if there is anything else we can clarify.

Thank you again for your time and for helping us improve the paper.

Best regards,

Authors

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

Similar to Pi3, the author studied another inductive bias in data preparison for feed forward reconstuction. The paper mainly study the influence of orientation of images in the field of feed forward reconstruction. The authors proposed a method ROVER. The method is training- free and use frozen DINOV2 features to select the right image orientation. It also comes up with a new datasets containing multi roll angles.

Critical strengths

The author evaluate many datasets and models, showing the effectivess of the method and the dataset. In addition, the method is training free, requires no reconstuction model retraining and can be applied to different downstream tasks.

Critical weaknesses

The paper use DINOV2 to study the influence of orientaion angles and to select the corrected angle, the paper define a score, which computes disparity of different patches. The authors are encourage the justity the selection of the score by statistical experiments. What cases this score can and can't describe the correct orientation. As this is still inductive biases (defined by human), I encourage the author to give detailed explaination why this method work. In the experiment part, the rotated image are abtained: image-> 3D GS -> rotation rendering ->optional diffusion. So the observed performance can be influence by roation or 3D rendering. What's the influence of the rendering quailty to the final results.

Questions for authors

See the weakness part. please provide more evidence to prove the effective of the methd.

4 comments
AuthorsOpenReview ↗

Statistical Experiments Added, Influence of the rendering quailty discussed

Selection of the scores by statistical experiments

We provided two experiments to justify the use of the scoring function, and following the reviewer's suggestion, we added a third experiment. First, the ablations in Table 5 compare DINOv2 against alternatives including DINOv1, homography, GeoCalib, and RotNet-based methods, demonstrating the effectiveness of DINOv2 representations for orientation estimation. Second, we conduct a systematic statistical analysis of spatial variance across rotation sweeps from 0° to 360° in 15° increments, repeated across four datasets, demonstrating that it consistently attains its maximum at the upright orientation (Supp. Section C). Finally, to isolate whether spatial variance is particularly effective at capturing this dependency, and in line with the reviewer's suggestion, we additionally compare it against alternative selection of scores, spatial covariance and spatial entropy. These experiments help us conclude that spatial variance of DINOv2 embeddings provides the most reliable estimate of canonical orientation among the evaluated statistics.

Orientation scoreMean angular error (°) ↓Accuracy within 5° ↑Accuracy within 10° ↑
Spatial variance (ours)10.355.7%56.9%
Spatial entropy63.85.4%5.4%
Spatial covariance35.141.7%41.7%

Score Analysis

Our score is particularly suitable for scenes where the spatial variance of visual features provides a consistent orientation-dependent signal (L 234-250). This is motivated by the fact that DINOv2 embeddings are trained with the KoLeo regularizer. As part of our evaluation, we also report an isolated failure case in the ETH3D office scene (Supp. Section I). In this scene, ROVER-based correction degrades all reported pose and point-cloud metrics across every evaluated model. We attribute this behavior to the scene's particular geometric and viewpoint characteristics. But, the reported failure mode is rare, occurring in only one of the 13 ETH3D scenes, and we observe no comparable degradation on any of the other evaluated datasets. Despite the occasional failure, ROVER achieves the lowest orientation error among all evaluated alternatives, with a mean error of 10.3° on NRGB-D, compared to 39.1° for RotNet and 56.3° for GeoCalib (Table 5). The same ordering is observed on 7-Scenes and Tanks & Temples (Table G).

Influence of the rendering quality

We account for the influence of rendering quality through a (1) multi-stage quality-control pipeline and (2) an additional quantitative validation experiment introduced in response to the reviewer’s suggestion. We perform extensive quality control (Supp. Section H) to minimize rendering artifacts. First, we fit a state-of-the-art 3DGS model to the scene, enabling novel-view synthesis. We then optionally apply the DifiX3D+ (Figures J–K), followed by manual inspection and filtering of the generated views (Figure L). This process ensures that the evaluated synthetic images are of high visual quality and minimizes potential confounding effects from rendering artifacts. To isolate the effect of rendering quality, we conduct an additional experiment in which we synthesize upright (0) images using Controlled Rotation Synthesis, retaining the original camera poses from which the images were captured (Supp. Section N; Table R). We then evaluate downstream performance against the original images to determine whether the rendering process itself affects reconstruction quality, independently of camera rotation. The results show that the rendering process introduces negligible changes in downstream performance, with differences remaining within the observed variance across runs. This suggests that the rendering process itself does not materially affect reconstruction quality.

Summary

Taken together, these results show that UprightBias is a systematic failure mode of feed-forward reconstruction models, appearing across architectures, datasets, and real captures, and that ROVER mitigates it at inference time without retraining. We have addressed each of the reviewer's concerns with additional evidence and revised the paper accordingly.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer FTEg

With about a week remaining in the discussion period, we wanted to kindly invite you to take another look at our rebuttal and the updated results we have added based on your valuable feedback.

We have addressed all of your comments in detail, updated several figures and tables, added additional baselines (statistical score experiments, additional experiment for rendering quality influence in supplementary section N) and added score analysis. We would be very grateful if you could share any further thoughts, or let us know if there is anything else we can clarify.

Thank you again for your time and for helping us improve the paper.

Best regards,

Authors

The newly added zero-degree rendering experiment (Appendix N) is a useful control. However, showing that the rendering gap is negligible at 0° does not necessarily imply that it remains negligible at arbitrary rotation angles. Rotated camera views may expose regions with different reconstruction quality, potentially introducing angle-dependent artifacts that confound the observed orientation sensitivity.

AuthorsOpenReview ↗

Results for redered rotated view included

We thank the reviewer for their response

To test whether rendering artifacts at rotated viewpoints confound our results, we applied Controlled Rotation Synthesis to the upright RotMV images and compared the rendered rotated views with the real captured rotated views of the same scenes and angles. Rendered views do not degrade performance more than real captures (results in table below), so rendering artifacts do not account for the sensitivity to orientation.

ModelN.C. ↑ capturedN.C. ↑ renderedAUC$_{30}$ ↑ capturedAUC$_{30}$ ↑ rendered
VGGT-Ω0.7480.77094.699.7
π³0.8060.79993.295.9
MapA0.7410.76188.490.6
DA30.7600.74194.199.5

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief Summary of Paper Revisions and Responses

We sincerely thank all reviewers for their thoughtful and constructive feedback. The revised manuscript incorporates the following updates to address the raised concerns

∙\bullet Table 6 caption revised to neutrally describe the reconstruction performance comparison between the baseline and ROVER-integrated pipelines.

∙\bullet Direct Rotation-Angle Regression through MLP Head (Supp. Section M): Additional experiment of a lightweight MLP for direct rotation-angle prediction as an alternative to ROVER's variance-based estimation.

∙\bullet Validation of Controlled Rotation Synthesis Fidelity (Supp. Section N): Compares reconstructions from original images against those processed through the Controlled Rotation Synthesis pipeline with identity rotation, verifying that the pipeline does not introduce systematic geometric distortions or reconstruction degradation.

∙\bullet Validation of Pose Estimation Robustness for RotMV (Supp. Section O): Additional evaluation that compares VGGT-derived camera poses against independently estimated COLMAP poses, demonstrating that ROVER's downstream reconstruction performance remains consistent across the benchmark scenes.

∙\bullet Alternative Scoring Functions for the Rotation Amount Prediction (Supp. Section P): Evaluates spatial entropy and spatial covariance as alternatives to spatial variance for orientation estimation on NRGB-D, demonstrating the superior performance of ROVER's variance-based scoring function.

All changes introduced in the main paper and supplementary are highlighted in blue text for easy identification.

We hope the revised paper can effectively address reviewer's concerns. Please let us know if there are any further suggestions.

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

Related papers

Discussion (0)

Sign in to comment.