acceptodds
Under review as a conference paper at ICLR 2027

UniMetric: Universal Metric Scale Recovery for Visual Geometry Foundation Models

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. To evaluate metric accuracy 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. Our codebase and the MetricBench benchmark will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Visual Geometry Foundation Models, Metric Scale Recovery, 3D Reconstruction

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_z38SOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Visual geometry foundation models recover structure and camera motion in an arbitrary unit. UniMetric estimates one positive scale for a sequence without training or extra sensors. A vision–language model (Seed1.8) proposes familiar objects, height intervals, confidences, and a support surface; SAM 2.1 produces masks. Object height is measured in the reconstruction along the normal of a plane fit to the support, and each height interval becomes an interval on the shared scale. A convex weighted sum of squared distances to these intervals is minimized by golden-section search, and the resulting scalar multiplies point maps and camera translations. MetricBench evaluates this correction on twelve driving, roadside, and embodied subsets (694 sequences) without ground-truth scale alignment. Across nine models, the authors report mean reductions of 54.2% in depth Abs Rel, 83.8% in Chamfer distance, and 65.6% in trajectory error, with every model improving on all three tasks at the suite level.

Critical strengths

1.The evaluation asks the question that standard GFM tables remove. Depth and pose are scored after rigid alignment only, never after fitting scale to the sensor reference. Under that protocol, most driving results start at δ₁ = 0.0 and Abs Rel near 0.97. Suite-level Abs Rel then falls from 0.892 to 0.338 on driving and from 0.671 to 0.379 on embodied scenes, and the same scale improves Chamfer distance and ATE for every model in Table 3. The gain is not confined to models with an arbitrary unit: π³ already has driving Abs Rel 0.330 and δ₁ 33.0, and UniMetric moves these to 0.190 and 80.6. Holding the GFM output fixed and changing only the scale makes the comparison easy to interpret. 2.Where a practitioner’s alternative is reported, UniMetric is ahead, and the design ablation matches the paper’s claim. On VGGT-Ω, one scale taken from Depth Pro, UniDepth v2, Depth Anything V2, or ZoeDepth is worse than UniMetric on Abs Rel, δ₁, Chamfer distance, and ATE for both nuScenes and FusionPortable (Table 4). Replacing the joint intervals by the single lowest-weight anchor drops δ₁ from 82.0 to 45.4 on nuScenes and from 52.2 to 18.0 on FusionPortable (Table 5). That is evidence that aggregating several uncertain sizes matters on these sequences.

Critical weaknesses

1.The argument that uncertain semantic intervals are what produces a consistent scale is not isolated. Proposition 1 is the mean-square bound for the sample mean under zero-mean, uncorrelated errors and exact relative heights. The controlled study uses those exact heights, setting hᵢ = q Hᵢ with q = 0.3, so it does not include reconstruction error. In the actual system, repeated frames of the same object enter the sum as separate observations, and category weights α_c appear in Eq. 9 without a definition or a statement of how they were chosen. There is also no comparison to a detector with a fixed nominal height per class, which is the classical version of this idea. Until that baseline exists, the VLM intervals and the convex objective are not shown to be necessary.

2.The headline reductions describe a correction of a bad global unit more than a uniform metric reconstruction, and the alternative is only tested narrowly. Driving δ₁ of 0.0 before correction means almost any plausible scale produces a large relative drop. After correction, VGGT remains at Abs Rel 0.722 on KITTI and 0.866 on DAIR-V2X-I, Scal3R at 0.856 on DAIR, and π³ on KITTI moves only from 0.190 to 0.188. Table 4, the comparison against metric-depth scale donors, covers one model and two of the twelve subsets. The oracle gap on those two subsets is still material (nuScenes Abs Rel 0.152 vs. 0.087; FusionPortable 0.331 vs. 0.219).

Questions for authors

1.On the full MetricBench protocol, how does UniMetric compare with a fixed nominal height per category, using the same masks and support normals, and with the same pipeline driven by a second vision–language model? If the nominal-height variant matches Table 3, I would treat the method as classical object-height scaling and lower the rating.

2.Please define α_c and state whether any of its values, the observation filters, or the choice of partial subsets (Humanoid Everyday and 3EED) were selected using MetricBench numbers. Also state which point inside a nonempty interval intersection is returned, since Eq. 11 is flat on that intersection.

3.Please extend Table 4 to the subsets where the corrected error stays high, at least KITTI and DAIR-V2X-I, for VGGT and π³. If a metric-depth scale is better on those subsets, the suite-level claim should be qualified.

2 comments
AuthorsOpenReview ↗

Response to Reviewer z38S

We thank Reviewer z38S for the precise comments (revisions in blue).

W1 / Q1. On the full MetricBench protocol, how does UniMetric compare with a fixed nominal height per category and with a second VLM?

Thank you for stating this test so clearly. Both comparisons are now in the ablation study of the main text (Sec. 4.4, Tables 5 and 6). The baseline replaces the VLM intervals with fixed category intervals taken from each dataset's own statistics and keeps everything else the same, including the masks and support normals. It is run with VGGT-Ω on one subset from each MetricBench suite, nuScenes and FusionPortable:

nuScenesFusionPortable
VariantAbs Rel↓δ₁↑CD↓ATE↓Abs Rel↓δ₁↑CD↓ATE↓
Fixed category intervals0.18363.21.4636.8640.37442.40.7848.331
VLM point priors0.17279.10.9895.2310.45232.41.80619.430
UniMetric, Qwen3-VL0.16270.80.7942.7540.32258.50.6214.002
UniMetric, Seed1.80.15282.00.7262.4760.33152.20.6534.238

The fixed-prior variant does not match UniMetric on either subset. With Qwen3-VL as the second VLM, UniMetric stays above the fixed-prior variant on both subsets.

W1. Proposition 1 and the controlled study assume exact relative heights and uncorrelated errors.

Thank you for pointing this out. You are right that Proposition 1 and the controlled study rest on these assumptions. The MetricBench experiments rely on neither. Every MetricBench result uses heights measured on the GFM reconstruction and counts repeated views separately, and Abs Rel improves in all 108 model–subset pairs. The analysis also extends. If M distinct objects are each seen m times with same-object correlation ρ, the mean-square error is bounded by σ²[1+(m−1)ρ]/(Mm), which still decreases as 1/M. If a measured height is off by a relative error η, the true scale stays inside the anchor's interval as long as H(1+η) stays within the VLM interval.

Q2. How is α_c defined, were its values, the filters or the partial subsets selected using MetricBench numbers, and which point of the intersection is returned?

Thank you for the question. α_c is a predefined category-level weight that reflects how reliable a category is as a scale anchor. Selected categories, such as cars, buses, doors and people, get their own weights from domain knowledge, and all others get the default (Eq. 9, Appendix C.2). The weights were fixed a priori, not tuned on MetricBench. For Humanoid Everyday and the quadruped part of 3EED, we removed repeated recordings of the same task in the same environment and kept 30 of 45 and 8 of 13 sequences. Redundancy was the only criterion. The minimizer is unique unless the intervals share an intersection of positive width (Proposition 2), where every scale has zero objective value. Our golden-section search keeps the right-hand bracket on ties, so it returns the upper end of the intersection.

W2 / Q3. Extend Table 4 to KITTI and DAIR-V2X-I for VGGT and π³; the headline mostly reflects the correction of a bad global unit.

Thank you for this suggestion. Table 4 now covers VGGT, VGGT-Ω and π³ on five subsets, including KITTI and DAIR-V2X-I, against seven other scale sources, with a new paragraph in Sec. 4.3. Abs Rel is below (lower is better, best in bold; Oracle fits the scale to ground truth).

GFMDatasetDepth ProUniDepth v2DA-v2ZoeDepth-NKFreePoseRSAVGLDUniMetricOracle
VGGT-ΩnuScenes0.2120.2780.3470.2180.7450.8670.6340.1520.087
KITTI0.4680.3800.4700.3670.9150.3920.3840.3470.122
DAIR-V2X-I0.6750.6150.6680.8800.9060.9210.8820.5420.132
FusionPortable0.5430.4850.3960.4320.8610.8740.8280.3310.219
JRDB0.2790.3640.8670.7360.4720.7320.7160.3120.227
VGGTnuScenes0.2130.3320.4100.2930.7720.5880.4450.3160.124
KITTI0.7530.7380.8250.6540.7770.7240.6780.7220.161
DAIR-V2X-I0.7690.7170.7660.8770.9870.8420.8010.8660.137
FusionPortable0.8670.8240.8360.8640.8280.7110.7090.6760.477
JRDB0.5290.6480.8780.6660.6200.9920.9220.5200.321
π³nuScenes0.2000.0960.3630.2130.7520.9470.8850.0980.069
KITTI0.3080.1980.2870.1940.7780.2230.2170.1880.082
DAIR-V2X-I0.6780.6120.6690.8210.8310.8690.8370.4710.142
FusionPortable0.4980.2730.3410.4710.8400.9220.9140.2190.189
JRDB0.2660.3450.4820.7150.5900.8420.8310.2530.208

Across the 15 settings, UniMetric is the most accurate in 10 and the second most accurate in two more. You are right that the gain over unscaled outputs mainly comes from fixing the global unit, and Table 4 compares UniMetric with other ways of fixing it.

AuthorsOpenReview ↗

Follow-up response, summary of manuscript revisions

To complement our response, we list below the changes in the revised paper that answer your comments. They are marked in blue, and we hope you will have a chance to look at them.

  • Added the fixed-prior baseline (Section 4.4, Table 6): replaced the VLM intervals with fixed category intervals, keeping the same masks and support normals.
  • Added a second VLM (Section 4.4, Table 5): ran UniMetric with the open-weight Qwen3-VL.
  • Extended Table 4 (Section 4.3; Appendix E.1): it now covers VGGT, VGGT-Ω and π³ on five subsets, including KITTI and DAIR-V2X-I.
Reviewer_zsLQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniMetric estimates a single metric scale for the output of a visual geometry model without training an additional scale-prediction network. A vision-language model supplies plausible object-height intervals and confidence scores, while segmentation and reconstructed geometry provide relative object heights. A weighted interval objective combines observations across objects and frames, and the resulting scale is applied consistently to depth, point clouds, and camera translations. Evaluation spans nine geometry models and twelve dataset subsets, with additional comparisons between scale estimators applied to the same reconstruction. The breadth of the evaluation and the competitive calibration results support practical value and a weak acceptance recommendation. The main limitations concern the incremental contribution over simpler semantic scale estimators and reliability under errors in geometric measurement and object priors.

Critical strengths
  1. The method provides a broadly applicable way to calibrate existing reconstructions. A single estimated factor makes depth, point clouds, and camera translations metrically consistent without retraining the geometry model. Evaluation across nine models, twelve subsets, 694 sequences, and 51,339 frames gives substantial evidence that the approach is useful beyond one backbone or scene type. The evaluation preserves scale errors rather than fitting scale to ground truth, so it directly measures the intended capability. This practical applicability is a principal reason for my positive assessment.

  2. The strongest comparison demonstrates benefits over alternative calibration sources on the same geometry. Table 4 fixes VGGT-Ω reconstructions and restricts each method to one scale per sequence. UniMetric achieves depth AbsRel of 0.152 on nuScenes and 0.331 on FusionPortable, compared with 0.212 and 0.396 for the best alternative calibrators in the table. This controls the underlying reconstruction and supports the value of semantic object anchors beyond simply correcting an arbitrary-scale output.

Critical weaknesses
  1. The contribution of the particular prior and aggregation scheme is not fully established against the closest alternatives. Object-height priors have previously been used for global scale recovery, including Sucar and Hayet, CVPR Workshops 2017, and language-based scale estimation is explored in the cited RSA and VGLD. UniMetric's specific contribution lies in combining open-vocabulary intervals, geometric measurements, and a sequence-level estimate. Comparisons using fixed category-height priors or robust aggregation of the same object observations would help determine which components add value. Table 5's single-anchor baseline selects the lowest-weight observation, making it a weak test of whether many observations and interval aggregation are necessary. A highest-confidence anchor and a robust multi-anchor estimator would be more informative. This limits the algorithmic novelty established by the current evidence, even though the overall integration remains useful.

  2. Reliability under errors in the measured geometry and correlated observations remains uncertain. Partial visibility, outlying reconstructed points, and support-plane errors can bias the max–min height measurement. An underestimated relative height produces an overestimated scene scale, potentially affecting every depth value and camera translation. The point-prior analysis assumes exact relative heights and zero-mean, uncorrelated prior errors; Figure 2 also constructs relative heights from ground-truth annotations. These results therefore do not characterize the principal geometric measurement errors in the full pipeline. Repeated observations of the same object or systematic category-size errors may remain correlated across frames. The paper needs clearer evidence on failure rates, large-error cases, and behavior when no reliable anchor exists. These limitations qualify the claimed generality.

Questions for authors
  1. Method contribution: With detection, segmentation, and geometry held fixed, how do category-height priors with robust aggregation, the highest-confidence single anchor, and relevant language-based estimators compare? This would determine how much value comes specifically from the VLM intervals and their aggregation.

  2. Reliability: How are repeated objects across frames handled, and what happens when no usable anchor is found? What are the failure rate and upper-tail scale errors when heights are measured from the actual reconstruction, including sequences where additional observations worsen the estimate?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zsLQ

We thank Reviewer zsLQ for the encouraging and concrete review (revisions in blue).

W1 / Method contribution. On fixed geometry, how do category-height priors with robust aggregation, the highest-confidence single anchor, and language-based estimators compare?

Thank you for these suggestions. We added them to the ablation study in Sec. 4.4 (Table 6), where each row changes one component of the VGGT-Ω pipeline:

nuScenesFusionPortable
VariantAbs Rel↓δ₁↑CD↓ATE↓Abs Rel↓δ₁↑CD↓ATE↓
(a) Fixed category intervals0.18363.21.4636.8640.37442.40.7848.331
(b) Single anchor (max)0.31748.61.86817.2550.64821.35.55145.548
(b) Single anchor (min)0.33445.41.98321.6430.68318.06.08648.649
(c) Point objective0.17279.10.9895.2310.45232.41.80619.430
(c) Median-ratio0.21958.61.6328.9880.38539.20.81510.300
(c) Gaussian MAP0.19368.51.1267.7300.36745.00.7216.699
UniMetric (ours)0.15282.00.7262.4760.33152.20.6534.238

The fixed category intervals, the counterpart of classical category-level priors (Sucar and Hayet; Sec. 2), are taken from each dataset's own statistics, yet they are worse than VLM intervals on every metric. The highest-confidence single anchor you asked for remains far from the multi-anchor result, and the interval objective is the best aggregation rule on every metric. For the language-based estimators, we added RSA and VGLD to Table 4 (Sec. 4.3), together with the object-size estimator of FreePose. Abs Rel is below (lower is better, best in bold; Oracle fits the scale to ground truth).

GFMDatasetDepth ProUniDepth v2DA-v2ZoeDepth-NKFreePoseRSAVGLDUniMetricOracle
VGGT-ΩnuScenes0.2120.2780.3470.2180.7450.8670.6340.1520.087
KITTI0.4680.3800.4700.3670.9150.3920.3840.3470.122
DAIR-V2X-I0.6750.6150.6680.8800.9060.9210.8820.5420.132
FusionPortable0.5430.4850.3960.4320.8610.8740.8280.3310.219
JRDB0.2790.3640.8670.7360.4720.7320.7160.3120.227
VGGTnuScenes0.2130.3320.4100.2930.7720.5880.4450.3160.124
KITTI0.7530.7380.8250.6540.7770.7240.6780.7220.161
DAIR-V2X-I0.7690.7170.7660.8770.9870.8420.8010.8660.137
FusionPortable0.8670.8240.8360.8640.8280.7110.7090.6760.477
JRDB0.5290.6480.8780.6660.6200.9920.9220.5200.321
π³nuScenes0.2000.0960.3630.2130.7520.9470.8850.0980.069
KITTI0.3080.1980.2870.1940.7780.2230.2170.1880.082
DAIR-V2X-I0.6780.6120.6690.8210.8310.8690.8370.4710.142
FusionPortable0.4980.2730.3410.4710.8400.9220.9140.2190.189
JRDB0.2660.3450.4820.7150.5900.8420.8310.2530.208

UniMetric is ahead of all three in 13 of the 15 settings and the most accurate overall in 10.

W2. Reliability under partial visibility, outlying points and support-plane errors remains uncertain.

Thank you for raising this. We agree that Proposition 1 and Figure 2 assume exact relative heights and independent errors. For actual reconstructions, where all these errors occur, we added a sequence-level analysis in Appendix E.5 with one new figure per metric (Figs. S4–S7). These figures show how Abs Rel, δ₁, CD and ATE change after scale recovery on every sequence, across all twelve subsets and nine GFMs: 6,246 model–sequence pairs. CD and ATE decrease on 92.5% and 84.2% of their pairs. For partial visibility, the VLM only selects objects whose top and bottom are both visible.

Reliability (a). How are repeated objects across frames handled, and what happens when no usable anchor is found?

Thank you for the question. Objects are not tracked across frames, so each sighting is a separate observation. The objective penalizes a scale only when it falls outside an interval (Eq. 11), so a repeated sighting adds weight only when the scale violates that object's interval. In the controlled study, the number of distinct objects grows only from about 11 to 18 between 20 and 50 observations, yet the mean scale error still falls on all four subsets (Table S1). When no usable anchor is found, UniMetric keeps the original GFM prediction (s = 1).

Reliability (b). What are the failure rate and upper-tail scale errors on actual reconstructions, including sequences where additional observations worsen the estimate?

Thank you for this question. Please see the new Appendix E.5, where Figs. S4–S5 show the change in Abs Rel and δ₁ for every model–sequence pair. On these actual reconstructions, Abs Rel improves on 88.1% of the pairs and worsens on 8.17%, and δ₁ improves on 91.9% and worsens on 3.94%. In the upper tail of Abs Rel, P99 rises from 0.997 to 2.828, while P90 falls from 0.987 to 0.560 and P95 from 0.991 to 0.854.

AuthorsOpenReview ↗

Follow-up response, summary of manuscript revisions

To complement our response, we list below the changes in the revised paper that answer your comments. They are marked in blue, and we hope you will have a chance to look at them.

  • Rewrote the ablation study (Section 4.4, Table 6): added fixed category priors, the highest-confidence single anchor and other aggregation rules.
  • Added language-based estimators (Section 4.3, Table 4): compared UniMetric with RSA and VGLD, as well as FreePose.
  • Added category-level priors to the related work (Section 2): discussed height-prior monocular SLAM and FreePose.
  • Added a sequence-level analysis (Appendix E.5, Figs. S4–S7): four new figures show how all 6,246 model–sequence pairs change after scale recovery.

Many thanks for your response. All of my concerns have been addressed.

I encourage the authors to include a discussion of reliability in the main paper. I also believe it is important to discuss the failure cases.

I will raise my score to 4.

Reviewer_N4X8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper estimates a single metric scale for the scale-ambiguous output of a visual geometry foundation model (GFM) using only RGB input. A VLM proposes objects, plausible height intervals, and a support surface. Then object masks are lifted into the GFM point map, and the resulting scale intervals are combined with a weighted loss. The estimate rescales both 3D points and camera translations. The paper also introduces MetricBench, a benchmark of 694 sequences across 12 subsets whose protocol does not fit scale to ground truth. Across nine GFMs, UniMetric broadly improves metric depth and trajectory estimates over unscaled outputs and outperforms metric-depth calibration baselines.

Critical strengths

S1. Simple, model-agnostic, and broadly evaluated. Across nine GFMs and 12 subsets, UniMetric improves Abs Rel in all 108 model–subset pairs.

S2. A useful scale-preserving benchmark. MetricBench evaluates depth, Chamfer distance, and ATE under rigid-only alignment across 12 cross-platform subsets, providing a useful benchmark for scale-ambiguous GFMs.

Critical weaknesses

W1. A relevant training-free precedent is omitted, and the competitive comparison is too narrow. Ponimatkin et al. (ICLR 2025, arXiv:2503.10307) retrieve typical sizes from a GPT-4-generated table, segment objects, measure their largest dimensions in a ZoeDepth point cloud, and take the median prior-to-measurement ratio as the scene scale. UniMetric remains distinct through intervals and multiple views, but this work should be discussed. The headline gains are also measured against unscaled outputs; on driving data, eight of nine GFMs begin with Abs Rel near 1 and essentially zero δ1, so large relative improvements mainly show that the estimate reaches the correct scale range. The stronger comparison in Table 4 covers only VGGT-Ω on two subsets. No prior median-ratio or Bayesian object-size estimator is evaluated, and Table 5 still uses the paper's own aggregation objective.

Questions for authors
  1. How often is the anchor set empty, and what scale is returned? Are such sequences excluded from the reported means under the "no valid data" rule?
  2. Height uses the full, unclipped range of projected points. Is confidence filtering applied per point, and how are visible-region measurements reconciled with whole-object height priors?
  3. The analysis assumes uncorrelated errors, but repeated sightings count separately, and performance in Table S1 improves little between 20 and 50 observations. How much of this plateau comes from sequences with fewer than 50 observations, and what happens when the count is based on distinct objects?
  4. What is the run-to-run variation from VLM sampling, how does an open-weight VLM perform, and what is the per-sequence cost?
2 comments
AuthorsOpenReview ↗

Response to Reviewer N4X8

We thank Reviewer N4X8 for the careful review (revisions in blue).

W1 (a). A training-free precedent is omitted, and the comparison is narrow and made against unscaled outputs.

Thank you for bringing FreePose to our attention. It also recovers scale from object sizes, within an object-pose pipeline, whereas we treat each size as an uncertain interval and let many objects across views jointly constrain a single scale. We added this discussion to Sec. 2 and FreePose to our baselines. We also extended Table 4 (Sec. 4.3) to three GFMs, five subsets and seven other scale sources. The Abs Rel results are below (lower is better, best in bold).

GFMDatasetDepth ProUniDepth v2DA-v2ZoeDepth-NKFreePoseRSAVGLDUniMetric
VGGT-ΩnuScenes0.2120.2780.3470.2180.7450.8670.6340.152
KITTI0.4680.3800.4700.3670.9150.3920.3840.347
DAIR-V2X-I0.6750.6150.6680.8800.9060.9210.8820.542
FusionPortable0.5430.4850.3960.4320.8610.8740.8280.331
JRDB0.2790.3640.8670.7360.4720.7320.7160.312
VGGTnuScenes0.2130.3320.4100.2930.7720.5880.4450.316
KITTI0.7530.7380.8250.6540.7770.7240.6780.722
DAIR-V2X-I0.7690.7170.7660.8770.9870.8420.8010.866
FusionPortable0.8670.8240.8360.8640.8280.7110.7090.676
JRDB0.5290.6480.8780.6660.6200.9920.9220.520
π³nuScenes0.2000.0960.3630.2130.7520.9470.8850.098
KITTI0.3080.1980.2870.1940.7780.2230.2170.188
DAIR-V2X-I0.6780.6120.6690.8210.8310.8690.8370.471
FusionPortable0.4980.2730.3410.4710.8400.9220.9140.219
JRDB0.2660.3450.4820.7150.5900.8420.8310.253

UniMetric is the most accurate in 10 of the 15 settings and ahead of FreePose, RSA and VGLD in 13.

W1 (b). No prior median-ratio or Bayesian object-size estimator is evaluated.

Thank you for this suggestion. We added Median-ratio, Gaussian MAP (a Bayesian estimator) and a point objective on interval midpoints to the ablation in Sec. 4.4 (Table 6(c)). All estimators use the same anchor observations from VGGT-Ω:

nuScenesFusionPortable
EstimatorAbs Rel↓δ₁↑CD↓ATE↓Abs Rel↓δ₁↑CD↓ATE↓
Median-ratio0.21958.61.6328.9880.38539.20.81510.300
Gaussian MAP0.19368.51.1267.7300.36745.00.7216.699
Point objective0.17279.10.9895.2310.45232.41.80619.430
Interval objective (ours)0.15282.00.7262.4760.33152.20.6534.238

The interval objective is the best on every metric.

Q1. How often is the anchor set empty, what scale is returned, and are such sequences excluded from the means?

Thank you for the question. The VLM finds candidate anchors in 687 of the 694 sequences (98.99%). If no valid anchor remains, UniMetric keeps the original GFM scale (s = 1). No sequence is excluded from the averages.

Q2. Is confidence filtering applied per point, and how are visible-region heights reconciled with whole-object priors?

Thank you for asking. The filtering is not per point. Confidence decides whether an anchor is kept (Appendix C.2), and all mask points of a kept anchor are used. The height is the max–min extent of these points along the support normal. The mismatch you describe is avoided when anchors are selected. The VLM only picks objects whose top and bottom are both visible, and it gives a height interval for that instance rather than a fixed category value.

Q3. How much of the plateau comes from sequences with fewer than 50 observations, and what if distinct objects are counted?

Thank you for this question. Only 13 of the 193 sequences (6.7%) have fewer than 50 observations. If we count distinct objects, 20 observations cover about 11 objects on average and 50 observations about 18. The last 30 observations add about seven new objects, and the rest are repeated views, whose errors are correlated. The rate in Proposition 1 assumes independent errors.

Q4. What are the run-to-run variation, the open-weight VLM performance, and the per-sequence cost?

Thank you for asking. We ran Seed1.8 and the open-weight Qwen3-VL-30B-A3B five times each with VGGT-Ω (Sec. 4.4, Table 5):

nuScenesFusionPortable
VLMAbs Rel↓δ₁↑CD↓ATE↓Tokens (M)Abs Rel↓δ₁↑CD↓ATE↓Tokens (M)
Seed1.80.153±0.0181.5±1.500.721±0.042.453±0.250.380.329±0.0251.4±1.500.649±0.034.138±0.423.81
Qwen3-VL0.162±0.0270.8±1.520.794±0.042.754±0.280.370.322±0.0258.5±1.580.621±0.044.002±0.392.89

Between runs, the standard deviation is 0.01–0.02 in Abs Rel and about 1.5 points in δ₁. Of the two VLMs, Seed1.8 is the more accurate on nuScenes, and Qwen3-VL is the more accurate on FusionPortable.

AuthorsOpenReview ↗

Follow-up response, summary of manuscript revisions

To complement our response, we list below the changes in the revised paper that answer your comments. They are marked in blue, and we hope you will have a chance to look at them.

  • Added FreePose to the related work (Section 2): discussed FreePose and earlier height-prior SLAM.
  • Extended Table 4 (Section 4.3; Appendix E.1): it now covers three GFMs and five subsets, with FreePose among the baselines.
  • Added median-ratio and Bayesian estimators (Section 4.4, Table 6): compared them with our interval objective.
  • Added a second VLM (Section 4.4, Table 5): ran Seed1.8 and the open-weight Qwen3-VL five times each, and reported the token cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.