acceptodds
Under review as a conference paper at ICLR 2027

Pose Shortcuts and Frame Reduction: When Do Canonical Views Help Spatial Verification?

Abstract

Rendering a relation-conditioned canonical view of a reconstructed scene is a natural way to make a spatial relation visually verifiable. We ask when such a view actually helps a verifier, using fixed-reference relations in 3D Gaussian splatting reconstructions of ScanNet rooms, and find three answers. First, a canonical camera is a function of the queried geometry: for left–right and front–back the camera rotation alone determines the room-frame label in 100% of our views, so a trained verifier that receives the pose need not read the image. A pose and-mask probe without RGB reaches 99.6% and 91.1% on canonical views of ten previously unused rooms and inverts to 0.2% and 9.6% when only the rotation is replaced by the mirrored camera's; the ResNet verifier itself inverts from 66.2% to 33.9% under the same mirrored rotation, so its front–back training-by-view interaction ( points) is a pose effect. Second, an analytic ceiling shows that a mask-conditioned verifier built on canonical views cannot beat the noisy geometry that placed its camera except when the noise reverses the pair direction, whereas gravity-aligned cameras keep the up–down frame conversion exact; with 1 m center noise and correctly marked objects, GPT-4o verifying up–down in such cameras beats the noisy geometry by 9 points. Third, for vision–language models the value of a canonical view is frame reduction: asking the question in the image's own frame and converting the answer with the known camera axes makes canonical views beat default views on left–right for all four models we test, by 6 to 23 points (90.1% for GPT-4o), while the room-frame interface stays at 50–58% in every view. Left–right pairs a model tells apart in the image frame are almost always right; depending on the model, the canonical gain comes from fewer undecided pairs, an exact frame conversion, or both. Canonical views help when they turn a frame-composition problem into a perception problem the verifier can solve; reported gains must be checked against pose shortcuts and geometry ceilings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
spatial reasoning, viewpoint selection, vision-language models, 3D Gaussian splatting, compositional reasoning, generalization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_peygOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents diagnostic results on how canonical camera pose encodes the queried direction, and estimated geometry error can propagate through camera placement to the verifier, and for VLMs, canonical view reduces the spatial reasoning problem to a perception problem.

Critical strengths
  1. This paper provides a quite interesting angle on how camera pose can be a shortcut utilized by verifiers that receive camera pose and investigates when canonical viewpoints can genuinely improve spatial reasoning.
  2. The paper combines systematic evaluations to investigate the effects of viewpoint selection and coordinate-frame transformations.
Critical weaknesses
  1. The proposed study assumes access to rerendered 3D scenes, and access to query-conditioned canonical views, however, this is not realistic in real-world spatial reasoning tasks. Accurate 3D reconstruction and camera placement are not practical. Realistic spatial understanding requires the model to take fixed or naturally shifted viewpoint images.
  2. Task diversity is limited. Given how Front–back is at 47–54% for every model, frame and view. The question remains whether the conclusion stands for more diverse spatial relationships and realistic spatial understanding tasks.
Questions for authors

The proposed approach relies on reconstructed 3D scenes and the ability to render query-conditioned canonical views. How does the existing experiments support the practical applicability of the proposed frame-reduction approach beyond the controlled evaluation setting? Minor: Please tighten the paper writing. The paper's writing is somewhat fragmented and reads as a collection of observations and experimental findings rather than a coherent paper narrative.

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

This paper investigates when canonical views help spatial relation verification in 3D reconstructed scenes. The authors show that camera poses can leak spatial relation labels, allowing trained verifiers to exploit shortcuts without relying on visual information. They also analyze the limitations of visual verification under noisy geometry and find that canonical views improve left–right reasoning when VLMs predict relations in the image frame and coordinate conversion is handled externally. These findings are supported by experiments on ScanNet scenes using ResNet-based verifiers and several VLMs.

Critical strengths
  1. The paper identifies a potential shortcut in query-conditioned camera selection, where camera parameters can implicitly encode the spatial relation being evaluated.
  2. The paper examines canonical-view effects from several perspectives, including trained verifiers, geometry-based analysis, and VLM prompting.
Critical weaknesses
  1. The paper relies on annotation-derived 3D object centers to construct canonical cameras and define ground-truth spatial relations. However, given these coordinates, the clean spatial verification task can already be solved directly through geometric comparisons. Although the authors acknowledge this and introduce a noisy-geometry setting, the additional experiments rely on synthetic Gaussian perturbations while retaining correctly identified 2D objects. This does not adequately capture the coupled errors in object grounding, geometry estimation, and camera selection that arise in realistic spatial reasoning systems. The practical value of the proposed analysis and its implications for real-world spatial verification remain insufficiently demonstrated.
  2. The experiments focus on three binary, axis-aligned spatial relations in a relatively small set of reconstructed indoor scenes, without evaluation on established spatial reasoning benchmarks. Moreover, the VLM improvements are primarily concentrated on left–right relations, while front–back performance remains close to chance and up–down results are inconsistent across models. This raises questions about how much of the observed benefit reflects a general advantage of canonical views versus the limitations of the particular prompting interface and the use of explicit geometric conversion.
Questions for authors
  1. Could the authors evaluate the findings on more challenging spatial reasoning settings, such as established spatial QA benchmarks, non-axis-aligned relations, or scenarios involving imperfect object grounding?
  2. Could the authors provide evidence using realistically estimated object coordinates and grounding results, rather than only synthetic coordinate noise and oracle 2D identification, to demonstrate when visual verification provides meaningful benefits over direct geometric reasoning?
Reviewer_YDV6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether choosing better camera views helps models judge spatial relations, using ScanNet scenes, ResNet models, and four VLMs. It reports that camera information can reveal the answer and that asking about image directions can improve accuracy. However, the manuscript repeatedly frames its contributions around correcting the authors’ earlier experimental protocols and conclusions, making it difficult to identify the main research question and contribution.

Critical strengths

None

Critical weaknesses

The paper does not establish a coherent research question or a convincing motivation for a standalone contribution. The introduction motivates the study through the statement that “earlier experiments of ours reported double-digit gains,” without identifying a corresponding publication or clearly explaining the status of those experiments. This dependence on an unspecified prior research history recurs throughout the manuscript: the contributions include “a corrected protocol and audit,” Section 7 refers to “our earlier interface,” and Appendix E states that historical numbers are “retained as motivation and audit targets.” These references leave the reader reconstructing the authors’ experimental history rather than understanding a clearly established scientific problem. Appendix E further acknowledges that the earlier results involved camera-dependent label distributions and omitted rendering failures. Correcting these problems is necessary, but the need to repair the authors’ own evaluation does not itself establish the significance of a new research contribution.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.