Pith.

Computer Vision and Pattern Recognition · arXiv:2610.16945 v1 · submitted 12 Oct 2026 · 14 references

UniMetric: Universal Metric Scale Recovery for Visual Geometry Foundation Models

Lei Yang, Yuner Zhang, Ziyang Liu, Ziying Song, Kai WU, Yiyang Gao, Jiangtao Su, Yiming Li

The paper says

“Recent advances in multimodal foundation models, exemplified by GPT-6 Astra, raise the prospect of general-purpose embodied reasoning and planning. Using Internet videos for embodied pretraining calls for reliable physical-scale annotations.”

more of the abstract

Yet visual geometry foundation models often recover scene structure and camera motion without reliable metric scale. We show that uncertain size priors of familiar objects can jointly constrain a reconstruction's shared scale. Building on this insight, we present UniMetric, a universal and training-free metric scale recovery framework for geometry foundation models, without additional sensors or task-specific metric supervision. Specifically, Metric Prior Acquisition elicits object height intervals from a vision-language model. Semantic-Geometric Alignment measures reconstructed object heights along the estimated support-surface normal in each frame and converts the priors into scale intervals. Global Scale Recovery reconciles these intervals through a convex objective, yielding a jointly feasible scale when they overlap and a unique compromise otherwise. The estimated sequence-level scale is applied uniformly to reconstructed points and camera translations. Controlled experiments with real VLM priors show that combining observations improves scale accuracy with exact relative object heights. To evaluate metric accuracy on GFM reconstructions without ground-truth scale alignment, we introduce MetricBench, covering twelve embodied, driving, and roadside subsets with sensor-derived ground truth. Across nine geometry foundation models, UniMetric reduces mean depth Abs Rel, reconstruction Chamfer distance, and trajectory error by 54.2%, 83.8%, and 65.6%, respectively, improving every model on all three tasks at the suite level. We release the UniMetric code and MetricBench data.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 14 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up. UniMetric is a competent, practically useful plug-and-play framework for recovering metric scale from VLM-provided object height priors, with thorough evaluation across many models and datasets. The idea is not entirely new but the execution is solid and the MetricBench benchmark is a real asset. This line is wrong

6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 14 references say.

6 minor objections

The paper says, at §3.1, controlled evaluation

Pith says: minor objection 1 of 6

The controlled study uses an arbitrary scale factor q=0.3 to construct relative heights. While this choice does not affect the relative improvements, the authors should explicitly state that the absolute scale error percentages scale with 1/q and that the conclusions about observation budget benefits are independent of this factor.

The paper says, at Table 4 and others

Pith says: minor objection 2 of 6

For the main quantitative results (depth, CD, ATE), no confidence intervals are provided. Given the bootstrap CIs reported in the controlled study (Table 1), it would strengthen the paper to include comparable uncertainty estimates for the main suite-level and per-subset results, especially to assess the variability across sequences within each subset.

show 4 more

The paper says, at §2, related work

Pith says: minor objection 3 of 6

The discussion of metric depth methods (e.g., Depth Pro, UniDepth) could be expanded to clarify that those methods predict metric depth per image, whereas UniMetric recovers a single scale for a whole sequence. This difference in granularity is important for interpreting Table 7.

The paper says, at §C.1, Prompt design

Pith says: minor objection 4 of 6

The prompt templates (Figures S2, S3) are detailed and well-motivated. However, the category reliability weights (α_c) used in Eq. (9) are given fixed values (0.95, 0.85, 0.70, 0.55) without justification for the specific numbers. A brief explanation or reference for how these values were chosen would improve reproducibility.

The paper says, at §E, supplementary results

Pith says: minor objection 5 of 6

The supplementary depth MAE/RMSE tables (S5, S6) and corresponding-point error tables (S7, S8) are useful but are not referenced in the main text. A sentence in §4.3 directing readers to these tables would help them interpret the additional metrics.

The paper says, at Table 3 and 4

Pith says: minor objection 6 of 6

Minor typo: in the abstract, 'reduces mean depth Abs Rel, reconstruction Chamfer distance, and trajectory error' – please ensure 'Chamfer' is spelled consistently (it appears correctly elsewhere).

What would settle it

Re-run the controlled study of Sec.3.1 but replace exact relative heights with heights measured from VGGT-Ω reconstructions on the same 193 sequences; if the MARE at K=50 exceeds the reported values by more than 3 percentage points, then GFM geometry errors amplify scale error beyond the VLM prior uncertainty alone.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 14 Oct 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 26 s · 40,493 tokens read, 558 written
  3. Editor wrote the verdict above 13 s · 40,835 tokens read, 1,015 written
  4. Referee accept: 6 minor 31 s · 41,262 tokens read, 854 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

You should know that this paper delivers exactly what it promises: a training-free way to add metric scale to any geometry foundation model by combining uncertain VLM height intervals with the model's own relative geometry. The empirical work is thorough nine foundation models, twelve dataset subsets, paired comparisons on depth, reconstruction, and trajectory. The reductions in Abs Rel (54%), CD (84%), and ATE (66%) at the suite level are clear and consistent.

What is actually new: the authors cast scale recovery as joint estimation from size intervals, formulate a convex interval objective, and use a VLM to provide open-vocabulary height priors. The controlled study in Section 3.1 separates semantic uncertainty from geometric errors, which is good practice. MetricBench itself is a useful cross-platform benchmark that avoids ground-truth scale fitting and will probably be reused.

The soft spots are minor and mostly acknowledged. The core idea object size priors for scale has been explored before (Sucar & Hayet 2017), but the extension to open-vocabulary VLM priors and the scale of the current evaluation make this a clear step forward, not a paradigm shift. The theoretical Proposition 1 assumes zero-mean uncorrelated errors with bounded variance, which is idealised; the authors treat it as motivation, not as a claim about real performance, so that is fine. They also state plainly that the method cannot fix local geometric distortions or trajectory drift, and that results depend on VLM and segmentation quality. I do not see any load-bearing flaw.

Who this is for: anyone working on turning Internet video into metric 3D data for embodied AI, or on scale recovery for monocular or multi-view geometry. It deserves a serious referee. I would bring it to reading group and would cite it if I were working in this area.

The strongest objection

The paper's central claim — that combining uncertain VLM height priors with GFM reconstructions improves metric scale across many models and datasets — is supported by extensive controlled experiments (Sec. 3.1), ablations (Table 8), and evaluations across nine GFMs and twelve datasets (Tables 3-6). The weakest assumption (VLM interval accuracy and GFM relative geometry consistency) is explicitly acknowledged and tested: the controlled study isolates VLM uncertainty with exact relative heights, and the full evaluation includes real GFM errors. The consistent improvement in all 108 depth pairs and 107 of 108 CD pairs, plus the suite-level gains across all three tasks, provides strong empirical evidence that the assumption holds on average. The paper's own limitation statements (Sec.5) correctly identify scope: dependency on semantic prior and object measurement quality, and inability to correct local distortions. I find no unresolved load-bearing concern that would undermine the central claim.

The reader's weakest assumption correctly identifies the key condition. The paper addresses it through multiple experiments. I concur that the evidence is sufficient to support the central claim, and no additional concern is needed. The structured output reflects this assessment without alteration to the reader's verdict.

The referee's report

If the results hold, UniMetric addresses an important practical problem: recovering metric scale for GFMs without additional sensors or metric supervision. The method is universal across models and scenes, and the improvements are substantial and consistent. The introduction of MetricBench provides a standardized evaluation protocol that retains scale error, which is valuable for future work. The paper demonstrates that uncertain semantic priors can jointly constrain scale, a plausible and promising direction.

Pith's reading of the paper (a machine's summary, not the authors')

The paper aims to establish that uncertain semantic size priors for familiar objects—provided by a vision-language model (VLM) as plausible height intervals—can jointly constrain the single metric scale of a scene reconstruction produced by any visual geometry foundation model (GFM). This is done without modifying the GFM, adding sensors, or training a new component. The authors introduce UniMetric, a training-free framework that measures reconstructed object heights along a support‑surface normal, converts each VLM‑supplied height interval into a scale interval, and then finds a single scale that best satisfies all intervals via a convex, piecewise‑quadratic objective. The paper reports that across nine GFMs on a new benchmark (MetricBench) covering driving, roadside, and embodied scenes, UniMetric reduces mean depth absolute relative error by 54.2%, Chamfer distance by 83.8%, and absolute trajectory error by 65.6%, improving every model on all three tasks at the suite level.

What carries the argument. The core mechanism is a convex, piecewise‑quadratic objective that sums squared distances from a candidate scale to the scale intervals derived from each VLM‑supplied object‑height prior and the corresponding reconstructed object height. The global scale is the minimizer of this objective, obtained by golden‑section search. When all intervals overlap, the minimizer is any point in the common intersection; when they do not, the unique minimizer balances the constraints. This method generalizes averaging point prior (the special case when each interval is a single point), and the authors prove mean‑square consistency for that case under mild assumptions.

Core claim. The central claim is that the metric scale of a GFM reconstruction—which is otherwise unknown up to a global factor—can be recovered by reconciling uncertain size intervals for familiar objects (cars, pedestrians, doors) obtained from a pretrained VLM. Because all objects in a scene share the same scale factor, the constraints from different objects intersect or overlap, narrowing the possible scale range. The paper formalizes this with a convex interval‑intersection objective and shows empirically that combining many such weak priors yields a scale that substantially improves depth, 3D geometry, and camera motion accuracy across a diverse set of GFMs.

The premise it rests on. The VLM's height intervals are accurate enough on average to meaningfully constrain the scale, and the GFM's reconstructed object heights are proportional to real‑world heights (i.e., the reconstruction's internal geometry is consistent up to an unknown global scale).

If the paper is right (Pith's inferences, not the authors')

UniMetric enables any existing GFM to produce metric‑scale outputs without retraining or hardware changes, making pretrained geometry models directly usable for applications that require real‑world units.

The approach can be applied to internet‑scale video, providing metric annotations for embodied‑Agent pretraining, as the VLM priors are obtained from RGB alone.

The MetricBench benchmark provides a standardized evaluation protocol for metric scale recovery, which may encourage further work on training‑free scale calibration.

Because the method separates scale recovery from geometry refinement, it can be combined with other post‑processing (e.g., trajectory smoothing) to further improve overall accuracy.

The method is not limited to heights; any VLM‑suppliable geometric dimension (width, depth) with a known prior and a measurable reconstructed extent could be used, potentially enriching the constraint set.

The convex formulation suggests that adding more objects (or frames) always improves or maintains the scale estimate, so the method can be made cost‑effective by adaptively selecting a small number of high‑confidence observations per scene.

The reliance on a VLM and SAM means domain shift (e.g., unusual object types or ambiguous support surfaces) could degrade performance; future work could incorporate multiple VLMs or on‑the‑fly prompt engineering.

The paper's limitation that a single scale cannot fix local geometric distortions or trajectory drift implies that for very long sequences, a per‑segment or per‑frame scale might be necessary, which UniMetric could straightforwardly be extended to handle.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Recent advances in multimodal foundation models, exemplified by GPT-6 Astra, raise the prospect of general-purpose embodied reasoning and planning. Using Internet videos for embodied pretraining calls for reliable physical-scale annotations. Yet visual geometry foundation models often recover scene structure and camera motion without reliable metric scale. We show that uncertain size priors of familiar objects can jointly constrain a reconstruction's shared scale. Building on this insight, we present UniMetric, a universal and training-free metric scale recovery framework for geometry foundation models, without additional sensors or task-specific metric supervision. Specifically, Metric Prior Acquisition elicits object height intervals from a vision-language model. Semantic-Geometric Alignment measures reconstructed object heights along the estimated support-surface normal in each frame and converts the priors into scale intervals. Global Scale Recovery reconciles these intervals through a convex objective, yielding a jointly feasible scale when they overlap and a unique compromise otherwise. The estimated sequence-level scale is applied uniformly to reconstructed points and camera translations. Controlled experiments with real VLM priors show that combining observations improves scale accuracy with exact relative object heights. To evaluate metric accuracy on GFM reconstructions without ground-truth scale alignment, we introduce MetricBench, covering twelve embodied, driving, and roadside subsets with sensor-derived ground truth. Across nine geometry foundation models, UniMetric reduces mean depth Abs Rel, reconstruction Chamfer distance, and trajectory error by 54.2%, 83.8%, and 65.6%, respectively, improving every model on all three tasks at the suite level. We release the UniMetric code and MetricBench data.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-14

The authors' main claim, as the classifier picked it out 0.96

“Across nine geometry foundation models, UniMetric reduces mean depth Abs Rel, reconstruction Chamfer distance, and trajectory error by 54.2%, 83.8%, and 65.6%, respectively, improving every model on all three tasks at the suite level.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.25 · human participants 0.03 · controlled comparison 0.50 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.59 · mentions code or data availability 0.99

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Recent advances in multimodal foundation models, exemplified by GPT-6 Astra, raise the prospect of general-purpose embodied reasoning and planning. background 1.00 0.04 0.02 hedged possibility 0.93
Using Internet videos for embodied pretraining calls for reliable physical-scale annotations. background 0.97 0.05 0.02 asserted 0.96
Yet visual geometry foundation models often recover scene structure and camera motion without reliable metric scale. gap 0.83 0.06 0.03 asserted 0.98
We show that uncertain size priors of familiar objects can jointly constrain a reconstruction's shared scale. contribution 0.54 0.76 0.21 asserted 0.91
Building on this insight, we present UniMetric, a universal and training-free metric scale recovery framework for geometry foundation models, without additional sensors or task-specific metric supervision. contribution 0.99 0.89 0.46 asserted 1.00
Specifically, Metric Prior Acquisition elicits object height intervals from a vision-language model. method 1.00 0.82 0.03 asserted 1.00
Semantic-Geometric Alignment measures reconstructed object heights along the estimated support-surface normal in each frame and converts the priors into scale intervals. method 1.00 0.90 0.04 asserted 1.00
Global Scale Recovery reconciles these intervals through a convex objective, yielding a jointly feasible scale when they overlap and a unique compromise otherwise. method 1.00 0.90 0.09 asserted 1.00
The estimated sequence-level scale is applied uniformly to reconstructed points and camera translations. method 1.00 0.80 0.05 asserted 1.00
Controlled experiments with real VLM priors show that combining observations improves scale accuracy with exact relative object heights. result 1.00 0.94 0.10 asserted 1.00
To evaluate metric accuracy on GFM reconstructions without ground-truth scale alignment, we introduce MetricBench, covering twelve embodied, driving, and roadside subsets with sensor-derived ground truth. contribution 0.78 0.95 0.04 asserted 0.99
Across nine geometry foundation models, UniMetric reduces mean depth Abs Rel, reconstruction Chamfer distance, and trajectory error by 54.2%, 83.8%, and 65.6%, respectively, improving every model on all three tasks at the suite level. result 1.00 0.96 0.71 asserted 1.00
We release the UniMetric code and MetricBench data. contribution 0.98 0.89 0.04 asserted 0.97

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 9,427 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (5)
Figure 1 of the paper
Figure 1. Uncertain object-size priors jointly constrain metric scale.
Figure 2 of the paper
Figure 2. Joint scale estimation with real VLM height priors and exact relative heights.
Figure 3 of the paper
Figure 3. Overview of UniMetric. Metric Prior Acquisition obtains object height intervals and masks from RGB images. Semantic–Geometric Alignment measures reconstructed heights along frame-specific support directions and converts the priors into scale intervals. Global Scale Recovery combines these intervals through a weighted convex objective, yielding a shared scale for reconstructed geometry and camera translations. Effect of the observation budget. Increasing the budget from K = 1 to 50 lowers MARE from 4.526– 7.245% to 1.086–2.175% across the four subsets (
Figure 4 of the paper
Figure 4. Qualitative scale recovery with VGGT-Ω on CEAR (top) and KITTI (bottom). Depth maps and camera trajectories compare predictions before (None) and after (UniMetric) scale correction. Point-cloud views show the highlighted regions and their measured heights alongside LiDAR references. These measurements cover visible regions rather than complete objects. GT denotes ground truth; rel. denotes the original reconstruction unit. 4.5 Qualitative Results We examine how the quantitative improvements appear in embodied and autonomous-driving scenes. For VGGT-Ω, UniMetric recovers scale…
Figure 5 of the paper
Figure 5. Camera trajectory scale recovery across eight GFMs.

Figures are the authors', reproduced from the paper.

References (14)
  1. [1] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, on PithPith's reviewarXiv
  2. [3] Model Scale CEAR JRDB FusionPortable DROID Humanoid Everyday 3EED RecoveryMAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓ VGGT None 6.9517.7767.059 8.507 37.490 42.829 0.521 0.672 0.438 0.592 14.534 17.992UniMetric 6.7407.9695.254 6.779 19.452 23.547 0.444 0.575 0.432 0... on Pith
  3. [4] Dongzhou Cheng, Taoran Yi, Ye Fang, Xingwu Zhang, Fan Feng, Yixuan Li, Gengxiong Zhuang, Rongze Wang, Shuai Yang, Wei Song, Weizhi Xue, Minyan Wu, Jie Gui, Jiaqi Wang, and Tong Wu. In-context robot learning with VLM agents.arXiv preprint arXiv:2609.19138, on PithPith's reviewarXiv
  4. [7] Diyari Mohammed Salih, Lingxiang Hu, Naima AitOufroukh-Mammar, and Fabien Bonardi. VIDAR: Visual-inertial dense alignment and reconstruction via a geometric foundation model.arXiv preprint arXiv:2607.17171, on PithPith's reviewarXiv
  5. [8] Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bo- janowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. VGGT-Ω.arXiv preprint arXiv:2605.15195, 2026a. Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlo... on PithPith's reviewarXiv
  6. [11] to produce one mask per region. Segmentation is performed independently on each image, so constructing the anchor observations does not require temporal mask propagation or persistent object identities. Anchor reliability and scale estimation.Candidate observations are screene... on Pith
  7. [12] without fine-tuning on MetricBench. For every frame of a sequence, the estimator predicts a metric depth map Dmdt from the RGB image It alone, without ground-truth intrinsics. The per-sequence scale is the median depth ratio over all frames, smd =median t,u∈Ω t Dmdt (u) Dt(u) ... on Pith
  8. [13] contains 45 Franka-manipulation sessions sampled from an external ZED view at 5 Hz, with depth-derived point clouds. For Humanoid Everyday (Zhao et al., 2025), we use 30 of the 45 G1 episodes (3,949 frames), sampled at 5 Hz from the 30 Hz source. For 3EED (Li et al., 2025), we... on Pith
  9. [2013] doi: 10.1177/0278364913491297. Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Kaixuan Wang, Hao Chen, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3Dv2: A versatile monocular geometric foundation model for zero-shot metric depth 14 UniMetric T echnical Report and sur... on Pithdoi:10.1177/0278364913491297
  10. [2022] Nikhil Keetha, Norman Müller, Johannes Schönberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, et al. MapAnything: Universal feed-forward metric 3D reconstruction.arXiv preprint arXiv:2509.13414, on PithPith's reviewarXiv
  11. [2023] Lei Yang, Xinyu Zhang, Jiaxin Yu, Jun Li, Tong Zhao, Li Wang, Yi Huang, Chuang Zhang, Hong Wang, and Yiming Li. MonoGAE: Roadside monocular 3D object detection with ground-aware embeddings.IEEE Transactions on Intelligent Transportation Systems, 25(11):17587–17601, 2024a. doi:... on Pithdoi:10.1109/tits.2024.3412759
  12. [2024] Zhenyu Zhao, Hongyi Jing, Xiawei Liu, Jiageng Mao, Abha Jha, Hanwen Yang, Rong Xue, Sergey Zakharov, Vitor Guizilini, and Yue Wang. Humanoid Everyday: A comprehensive robotic dataset for open-world humanoid manipulation.arXiv preprint arXiv:2510.08807, on PithPith's reviewarXiv
  13. [2025] Alessandro Burzio, Tobias Fischer, Sven Elflein, Qunjie Zhou, Riccardo de Lutio, Jiawei Ren, Jiahui Huang, Shengyu Huang, Marc Pollefeys, Laura Leal-Taixé, Zan Gojcic, and Haithem Turki. Déjà View: Looping transformers for multi-view 3D reconstruction.arXiv preprint arXiv:2605.30215, on PithPith's reviewarXiv
  14. [2026] ByteDance Seed. Seed1.8 model card: Towards generalized real-world agency.arXiv preprint arXiv:2603.20633, on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment