acceptodds
Withdrawn from ICLR 2027

From Prediction to Measurement: The Protocol Boundary for Material Parameters from Video

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video-based inverse-physics systems increasingly report material parameters in physical units, yet predicting physical motion and measuring physical parameters are different tasks. For the continuum models these systems use, jointly scaling density and every stress-dimensioned parameter leaves the observed mechanics unchanged, so the observations determine gauge-invariant combinations such as the elastic speed scale but not the absolute mass–stress scale. We turn this classical symmetry into a criterion for when an observation protocol supports an absolute measurement: the surviving gauge is exactly the subgroup that leaves invariant every exogenous datum the inference holds fixed. Kinematic and density-proportional forcing—passive motion, gravity, self-weight, and even displacement-controlled actuation—therefore preserve the gauge, while a fixed datum carrying an absolute mass or force scale removes it; with known geometry, a weighed mass suffices. The boundary governs evaluation as well. If every validation protocol preserves the gauge, every score computed from its observations is gauge-invariant, so held-out views, frames and kinematic interactions can validate predicted dynamics without validating the absolute parameters behind them: a candidate whose modulus is wrong by , with density co-scaled, still attains silhouette IoU , and one weighed mass exposes it. Across seven protocols, four constitutive families and released systems from two independent lineages, we find that when the mechanics leave the absolute scale undetermined it is supplied elsewhere—by an assumed density, a prior, an initialization or optimizer geometry—even where density is genuinely optimized. We state what to report, and how to validate, instead.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
video, physics, gauge-invariant

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2ATHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper addresses inverse-physics methods that predict material parameters from video inputs. Its main observation is that the absolute values of certain physical parameters cannot be inferred from observed motion alone. Specifically, the paper discusses elasticity and density, for which only their ratio can be inferred, rather than their absolute values. The paper argues that this limitation should be acknowledged in both the prediction and validation of video-based inverse-physics methods. It also investigates how existing methods nevertheless produce predictions of these absolute values.

Critical strengths

• The main motivation point raised in the paper is correct and clear: the community should consider whether the physical parameters predicted by video-based inverse-physics methods are actually identifiable from the video. • The paper investigates how absolute elasticity and density values are determined across several physics simulators.

Critical weaknesses

I find the paper to be very difficult to follow. I believe that improving this requires major revisions in the paper. The paper’s topic is outside my primary field of research, and I acknowledge that some of my criticism may stem from my limited familiarity with the field. Below are a few examples:

o Ideas are frequently expressed in overly complex and confusing language, even when they could be conveyed more simply. For example, “A datum that removes the symmetry must also couple to the realized motion for parameter values to be distinguishable” (line 215) could be simplified to “A datum that removes the gauge must have a noticeable effect on the motion observed in the video.”

o Another example of unnecessary complexity is Section 6. It seems straightforward that if density cannot be inferred from a video, then a predicted density value is not reliable, even if a density parameter is optimized within a physics simulator. I do not understand why an entire section is dedicated to this point. In my opinion, the sentence in lines 361–363 is sufficient to explain it.

o Many terms are not explained. For example:

  • “Boundary” repeats throughout but never explained. Same goes for “Classification”.

  • Lines 108-110: “gauge certificate” and “mass-scale transformations” are not explained. The sentence is unclear without defining these.

o Almost every paragraph begins with a short title, even in the Introduction, and many paragraphs are disconnected from those before and after them. This made the paper difficult to follow, as I often struggled to understand how the different paragraphs relate to one another and how the paper's argument develops.

o Some of the paragraph titles are confusing. For example, "The protocol determines the boundary" (line 78) is unclear, as I could not understand which protocol or boundary it refers to.

o Semicolons are overused improperly, resulting in long difficult-to-follow sentences. For example, line 113.

In addition, I believe the paper's contribution is too narrow:

o The main contribution is highlighting the inability of video-based inverse-physics methods to estimate the absolute values of elasticity and density. However, as far as I understand, this limitation is well-established in physics: the absolute values of elasticity and density cannot be inferred from video alone, and only their ratio can be determined (Feng et al., 2022), unless additional information is available. The implication for evaluation is immediate: evaluating the accuracy of absolute elasticity and density predictions means evaluating quantities that cannot be uniquely determined from the video.

The paper also investigates how different methods produce these absolute values, which is interesting. However, in my opinion, this analysis has limited practical value, as we know predictions cannot be reliable from the video alone.

Overall, I do not think the contribution is broad enough, as the paper mainly points to a known problem.

Questions for authors

I believe the paper could become stronger if it demonstrated how accounting for this limitation changes our understanding or evaluation of existing video-based inverse-physics methods. For example, would correcting the evaluation protocol change the ranking of existing methods? Could a method previously considered the best turn out to be inferior under a more appropriate evaluation?

Reviewer_Qd68OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

In the area of inverse physical-parameter estimation from video, the manuscript identifies a fundamental problem that has been widely overlooked: predicting motion and measuring physical parameters are two different things. Its central finding is that continuum mechanics possesses a scaling symmetry — when density and every stress-dimensioned parameter are co-scaled, the motion remains exactly unchanged. Passive video (including gravity, self-weight, and displacement-driven actuation) therefore determines only gauge-invariant combinations and can never determine the absolute modulus; only data carrying an absolute mass or force scale can break this gauge. The manuscript further proves that any evaluation metric computed from gauge-preserving observations cannot validate an absolute parameter. The authors audit released systems including PAC-NeRF, NeuMA, Vid2Sim, and M-PhyGs, finding that their reported absolute scales actually come from assumed densities, learned priors, or optimizer "selectors" rather than from the data itself; they also physically validate the symmetry with a real high-speed-video experiment using the aluminium–steel "matched-wave-speed pair." Finally, the paper proposes standardized guidelines for reporting and benchmark design.

Critical strengths

I think the biggest strength of the manuscript is that it identifies a fundamental problem that has been widely overlooked, drawing a clear distinction between "predicting dynamics" and "measuring parameters." Also,the benchmark-design implication (Corollary 2) is consequential and immediately actionable.

Critical weaknesses
  1. The manuscript treats the protocol as the variable and the audits, which is conceptual and critical rather than technical. No new estimator, algorithm, or benchmark dataset is introduced, the manuscript is closer to a well-executed analysis than a methodological advance.
  2. The protocol verification runs in the authors' own MLS-MPM solver - an objection they themselves flag. And the one experiment that would close the loop physically - the Appendix F anchor switch - was not executed.
Questions for authors

Given that the fixed rho = 1000 convention propagates through a shared code lineage, do you intend the audit as "one convention propagating" (as you state) or as evidence of prevalence in the field? How many additional independent systems would you need to audit to support the stronger claim?

Reviewer_8iXyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper is about video-based inverse physics methods that nowadays output numbers in physical units. The research question is: "when an observation protocol contains enough information for an absolute material parameter to count as a measurement".

The authors formalize which protocols can remove a given gauge, turning the limitation into a design rule - broad class of physics-informed pipelines can output an absolute quantity (e.g. elastic modulus) without that quantity being a mechanical measurement — because the absolute scale is fixed by a selector or prior, not by the observed mechanics. It then characterizes which protocols can remove a given gauge, and audits released systems to identify where the scale actually comes from, and demonstrate with real video how a declared anchor converts the invariant into an absolute estimate

The focus is on mass–stress gauge.

Critical strengths
  • useful conceptual framing; a method can output a number in physical units without that number being a mechanical measurement. The three-way split gives practitioners a concrete checklist to audit their own pipelines
  • The paper shows that prediction can be correct while the absolute parameter remains unvalidated, and that the boundary is protocol-dependent.
  • The paper reframes the question from "can video measure x?" to "what fixed the absolute scale?"
Critical weaknesses

The paper is more a position note that a contribution.

Questions for authors

I have some concerns with the relation of this paper with cited literature (see questions). In particular:

  • "Visual vibrometry: Estimating material properties from small motions in video" It seems to me the submitted paper is a generalization of that single case.
  • "On testing structural identifiability by a simple scaling method: Relying on scaling symmetries can be misleading" Does your gauge result fall into the "misleading" category that those authors describe?

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

Related papers

Discussion (0)

Sign in to comment.