acceptodds
Under review as a conference paper at ICLR 2027

When Does Geometry Add to Recognition? A Fusion Benchmark for Monocular Volume Estimation

Abstract

Monocular object volume estimation is judged by one error averaged over a dataset, mixing two contributions: (i) the recognised category, which already implies much of an object's volume, and (ii) the geometric pipeline that measures it. A benchmark that cannot separate them cannot say which did the work. We introduce OneVol, which factors a monocular estimate into six channels, scores each alone and in combination under rules, combiners, and an oracle on eight datasets, and first asks whether a dataset can support a metric claim. A measurement model gives a bar any estimator must clear, the ground truth's within-class spread, and five validity criteria: five of six volume datasets fail at least one, three on the data alone; and NOCS's 13,371 rows carry the evidence of 46 objects. The same model sizes the next benchmark: about 1/r² targets for a source of within-class correlation r, several hundred to resolve its gain, against the six volume datasets' 9 to 70 targets. The consequence: the median volume of same-class training objects beats the full pipeline in 23 of 24 dataset–backbone cells, and survives a predicted class, unresolved on MADIMA23 and MRGBD. On MetaFood3D, whose classes hold one object each so the prior is only a global median, the best published pipeline beats it. Fused on recognition's terms, the within-class regression keeps the prior and adds only the source's within-class deviation, lowering ECUSTFD's mean log error by 6–11% for every backbone, two of four surviving Holm's correction, and in 25 of 26 cells on the mass datasets, the only ones large enough, confirmed by a pre-registered test on held-out objects and dishes; a second registration, extending the rule to several sources, fails. A weak measurement should be judged by what it adds beyond a strong prior, not by its own accuracy. Code and predictions: https://github.com/nobody-eh/OneVol-review.

Primary area:
datasets and benchmarks
Keywords:
benchmark, evaluation validity, information fusion, monocular volume estimation, metric depth, vision-language models, pre-registration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FybLOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2(+1)
Read review
Paper summary

In the task of monocular food volume estimation, this paper uncovers a critical flaw in existing benchmarks: they conflate a model's geometric measurement capability with the contribution of simple category recognition. Furthermore, the authors demonstrate that the performance gains from popular VLM-fusion approaches do not actually arise from sophisticated visual reasoning by the VLM, but merely from pulling extreme overestimations by the geometric model back toward a common-sense scale. The paper also validates that accuracy can be improved by taking the class median as a reference and leveraging the geometric model solely to extract and calibrate within-class individual variations. Overall, these findings suggest that to measure the true contribution of geometric models, an "image-blind class average" must serve as a mandatory baseline, and that a rigorous evaluation protocol is necessary to rule out superficial improvements from naive VLM fusion.

Critical strengths

As multimodal fusion has become increasingly mainstream, uncovering the critical flaws and confounding factors in existing evaluation datasets is a timely and valuable contribution. It seems also impactful to demonstrate that VLM-fusion methods do not achieve genuine visual reasoning, but merely act as coarse scale regularizers that pull extreme geometric overestimations back toward common-sense priors.

Critical weaknesses

The manuscript suffers from poor writing and an overwhelming information density, requiring unreasonable effort from the reader to digest. Domain-specific acronyms such as NOCS, MADIMA, and ECUSTFD appear without any explanation or definition, disregarding readers outside this specific subfield. Figures and tables are frequently unreferenced or cited out of order, and the text leans far too heavily on a massive appendix that hardly feels manageable for a human reader. While the empirical results appear very important, the presentation fails to communicate them effectively, and a major overhaul of the manuscript's presentation is necessary.

Questions for authors

What do the authors consider to be the most essential contribution of this work? Currently, both the Abstract and the Introduction read like an enumeration of specific empirical results, making it difficult to grasp the overarching takeaway and fundamental implications for researchers in this field. I would like the authors to clearly articulate their perspective on this.

Reviewer_2NqNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces OneVol, a benchmark for monocular volume estimation that separates the contribution of object recognition from geometric measurement. The authors show that a simple class-based volume prior outperforms geometric pipelines across most evaluated settings. The paper further introduces benchmark validity criteria and evaluates the framework across multiple volume and mass datasets.

Critical strengths

The paper addresses an important and largely overlooked issue in monocular volume evaluation: strong performance may come from recognizing the object category rather than accurately recovering its geometry. The proposed evaluation framework is thorough, with careful controls for repeated observations, class priors, predicted labels, multiple geometric backbones, and statistical significance. I also appreciate the extensive validation and sensitivity analyses, including the pre-registered experiments on held-out data, which make the main empirical observations substantially more convincing.

Critical weaknesses
  1. The writing of the paper is unusually compressed and self-invented, making it hard for human readers to focus on the main contributions of the paper. for example, in the abstract, “A measurement model gives a bar any estimator must clear, the ground truth’s within-class spread…” or later on “A weak measurement should be judged by what it adds beyond a strong prior, not by its own accuracy.”. It is not that the sentences are not expressing the points in the paper; it is that the language is not very direct and is more complicated than needs to be to explain simple ideas.

  2. One weakness of the paper is that the volume variation can be captured by the dataset's ontology, essentially coupling volumetric estimation with classification. Random class merging demonstrates the dependence but does not establish robustness to meaningful alternative taxonomies. Experiments with semantically meaningful coarse/fine class hierarchies would strenghten the claims in the paper.

  3. The theoretical prediction of fusion gain is not strongly supported by the experimental evaluation. The paper reports an overall correlation of r=0.87r=0.87 between predicted and realized fusion gain, which sounds very strong. However, this is largely driven by the mass datasets: rmass = 0.99 rvolume = .45 and for the volume geometry cells specifically, the appendix reports only r=0.24r=0.24. Furthermore, the class-anchored rule actually hurts performance in 14/54 volume cells, whereas the population model itself cannot predict harm. The finite-sample condition catches 13/14 of those failures, which is a positive result, but this suggests that the theory is substantially better at identifying broad safe/unsafe regimes than at quantitatively predicting fusion gain. I think the paper should make this distinction much clearer.

Questions for authors

Not sure how much language refinement was done by LLMs but the paper requires significant cognitive effort to follow simple concepts. I am leaning towards weak acceptance because the motivation of the paper is strong but the language and presentation needs some work so the concepts can be communicated efficiently.

Reviewer_7QMYOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper argues that monocular volume benchmarks mix recognition (the class already implies the volume) with geometric measurement. A class-median prior beats the full geometric pipeline in 23 of 24 settings, most datasets fail basic validity criteria, and geometry helps only when fused as a within-class correction to the prior, confirmed by a pre-registered test.

Critical strengths

Important evaluation finding with very careful statistics (object-level bootstrap, constant-source control, pre-registration), and practical guidance on how large future benchmarks need to be.

Critical weaknesses
  • The positive result is small (about 2 cubic centimetres on a median 155 cubic centimetre object, as the authors acknowledge) and the fusion rule is standard within-class regression. The value of the paper lies in the evaluation, not in the method.
  • The four depth backbones share the same segmenter, scale rule and integration operator, so their agreement is closer to one test than four. The only fully valid volume dataset, MetaFood3D, is also where geometry wins, which weakens the headline claim.
  • The paper is very dense and many results needed to follow the argument are in a 94-page appendix.
Questions for authors
  • How should the main claim be read for a well-designed benchmark, given that geometry wins on MetaFood3D?
  • How robust are the dataset verdicts to the SPREAD threshold of 0.10? ECUSTFD fails at 0.098.
  • Would a pipeline with a different segmenter or scale cue change the conclusion?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.