acceptodds
Under review as a conference paper at ICLR 2027

EventGauge: When Can 3D Evaluation Validate Metric Scale?

Abstract

Learned 3D reconstruction increasingly claims metric depth and absolute scene scale, yet standard evaluation can remain unchanged when the recovered global scale is wrong. Similarity alignment can absorb global scale, while co-centred rendering can preserve projections under joint rescaling, removing scale evidence from both common reporting channels. We introduce EventGauge, a physical audit that withholds a calibrated displaced camera until reconstruction is frozen. The audit asks whether the finished reconstruction is best supported at the true physical viewpoint and whether that support is sharp enough to justify a metric conclusion. Controlled rescaling verifies that the response moves with scale as predicted by geometry. Across positive and failure cases, the audit can support physical consistency, expose a residual inconsistency, or remain unresolved when the evidence is too broad. EventGauge therefore turns a metric-scale claim from something standard reporting can silently absorb into a falsifiable physical test. Selecting a scale is a property of reconstruction; validating that scale is a property of evaluation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
metric scale validation, 3D reconstruction evaluation, scale ambiguity, physical-view evaluation, event-based vision

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

This paper introduce EventGauge a physical audit that withholds a calibrated displaced camera until reconstruction is frozen. The audit asks whether the finished reconstruction is best supported at the true physical viewpoint and whether that support is sharp enough to justify a metric conclusion. The method retaining trajectory-scale information while evaluating a frozen reconstruction from a withheld calibrated camera.
Two complementary checks are performed. The trajectory check keeps the reconstruction-to-metric conversion absorbed by similarity alignment and reports the residual when rescaling is forbidden. The image check then tests the frozen scene under that conversion using a displaced camera withheld from reconstruction and scale fitting. With EventGauge a sharp response can support a metric conclusion

Critical strengths

Proposing a 'aduit' is interesting and novel, I never read paper like this before.

The paper makes an interesting distinction between selecting a scale, producing scale-sensitive scores, and independently validating physical scale.

The prespective of withheld camera is insipiring.

The M3ED cases demonstrate why geometric parallax alone does not establish useful precision is also a valid point.

Critical weaknesses

This reviewer is confused about what is an 'audit'. Is this a benchmark? an error metric? or new dataset? Read through the paper, seems like 'audit' is a way of desiging a benchmark, but the authors do not actually provide a benchmark?

If I understand correctly, the 'audit' still need an oracle which is the metric scale offset information about the leave out camera. So this bascially means we have a matric scale reference, and the rest is how we should align to this reference?

Nevertheless, the paper is extremely difficult to follow. I think the authors may need more effort to better 'position' the paper and tell the read what is an 'audit'? It seems that the paper is talking about a new systematic way to measure the metric scale reconstruction? Is this a proposal for the follow up research? What is the problem of having other direct metric scale evaluation? because neverthelss it seems the method actuall still need a full metric scale dataset to test with?

Questions for authors

What is considered to be the techinal contribution of the paper?

What do you mean by 'scientific outcome' ? what is 'uncensored response' ?

Who should use EventGauge, or when it should be used? What is the difference than comparing the actual metric scale error metric?

How to determine a peak is sufficiently near one and sufficiently narrow to support a stated metric precision?

How to determine the reconstruction length unit in the IncEventGS comparison?

Reviewer_oM74OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper suggests that claiming the reconstruct a scene up to metric scale is not trivial and that a fully reconstructed scene might be correct in terms of geometry and rendering but still fail to reconstruct the scene by metric scale. It shows how to validate the metric-scale claims using a single held-out camera.

Critical strengths
  1. The claim of the paper regarding the decoupling between scene reconstruction and scale correctness is well justified.
  2. The problem of measuring the scale correctness is a real problem that worth investigating, solving such problem might be useful to the community.
Critical weaknesses
  1. Novelty and motivation – The novelty of the paper is not clear, the analysis of the fact that a 3D reconstruction is correct up to scale is known, and using a GT camera to prove the scale mismatch is not practical – If we have GT cameras wit intrinsics and extrinsics the reconstruction is straight-forward and we don’t need the scale prediction, the practical case is when there is no GT cameras at all.
  2. Missing evaluations – The paper suggests a set of check to validate metric scale, it is expected to show it on well known metric scale and non-metric scale baselines such as Depth-Anything, VGGT, Map-Anything and more. It is also hard to follow the paper since there are no qualitative examples, visualizations or benchmarks.
Questions for authors

None

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

This paper proposes EventGauge, an evaluation protocol for assessing metric scale in 3D reconstruction. The authors observe that Sim(3) trajectory alignment removes global scale errors. They propose evaluating a reconstruction from a calibrated displaced camera and sweeping the camera baseline. Using PSNR they can measure the location and sharpness of the image’s response, providing insight into scale errors.

Critical strengths
  1. The baseline sweep is an interesting extension of standard novel-view evaluation. Measuring the sharpness of the response may reveal information about the reconstruction that is not clear from evaluation at a single fixed pose.
Critical weaknesses
  1. The proposed metric’s meaning can be unclear, as it depends on geometry and appearance. The metric relies on image similarity, but this can be influenced by factors besides metric scale errors. For example, local geometry errors may produce an incorrect baseline for a reconstruction whose scale is accurate. The converse is also true: if the response curve is maximized near 1, this does not guarantee metric accuracy.
  2. The metric’s value is not well-established. The paper does not convincingly demonstrate that standard evaluation techniques are flawed. Typically Sim(3) alignment is done intentionally, but standard metrics such as novel-view rendering from a known metric camera pose (e.g. as in SHARP) can capture metric scale errors. This is mentioned in the paper, but the audit in section 3.5 instead focuses on papers that may not claim metric scale. It is unclear how existing evaluations are flawed and why a new metric is needed.
  3. The experiments do not demonstrate the usefulness of the proposed metric. The experiments in 5.1 and 5.2 simply confirm expected properties that are nearly guaranteed to hold. The remaining experiments do not demonstrate how this metric provides value in assessing the quality of 3D reconstructions.
  4. The writing is unnecessarily difficult to follow. The paper describes standard geometric facts in excessive detail, and does not motivate its own approach or the experiments very well.
Questions for authors
  1. In what situations would EventGauge provide insight not captured by SE(3)-aligned ATE or a novel-view synthesis evaluation metric from a specified metric camera pose?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.