- 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
-
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.
-
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.
-
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
- 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.
- 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
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.
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 → 0.73; DA3: 0.72 → 0.73) and changes in AUC₃₀ of at most 1.2% points (VGGT-Ω: 96.7 → 95.8; VGGT: 91.6 → 91.1; DA3: 97.4 → 96.5; π³: 90.7 → 90.1). MapAnything also exhibits negligible variation in N.C. (0.65 → 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.