acceptodds
Under review as a conference paper at ICLR 2027

CalibAdapter: Adapting 3D Foundation Models to Calibrated Cameras

Abstract

Recent feed-forward 3D foundation models have demonstrated strong capabilities in reconstructing scene geometry, supporting a wide range of downstream tasks. In practical settings such as robotic manipulation and autonomous driving, calibrated camera intrinsics and extrinsics are usually readily available through standard sensor-calibration pipelines. Existing works like MapAnything, DepthAnything3, and PromptDA have already shown that utilizing calibrated camera parameters can improve reconstruction accuracy. However, these methods do not guarantee that the predicted depth will back-project to geometry consistent with the provided calibration. We quantify this deployment discrepancy using **Calibrated Pointmap Error (CPE)**, which measures the resulting pixel-corresponded metric 3D error. To address this discrepancy, we propose **CalibAdapter**, a lightweight adaptation mechanism for incorporating known camera calibration into 3D foundation models. CalibAdapter attaches a projective sparse residual branch to a selected cross-view interaction block while keeping all pretrained parameters frozen. CalibAdapter consistently reduces CPE across **VGGT-Ω**, **VGGT**, **DA3**, and **π³**, achieving an average relative CPE reduction of **28.3%** together with overall improvements in conventional reconstruction metrics. It further yields gains on several held-out benchmarks. Beyond conventional reconstruction benchmarks, CalibAdapter improves geometric accuracy in **autonomous driving with known rig calibration and ego poses**, and increases **real-world robotic manipulation** success rate by 18.4 percentage points over the parameter-matched baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, camera calibration, parameter-efficient adaptation, autonomous driving, robotic manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_iKUhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The main motivation of the paper is that the cameras in many systems, including robotic and autonomous-driving systems are often calibrated or can be easily recalibrated, but the recent feedforward reconstruction models do not fully exploit this information. To remediate, the authors propose CalibAdapter, a lightweight framework that adapts such frozen 3D foundation models to exploit known camera intrinsics and extrinsics through an auxiliary residual branch. The model combines projective camera transformations inspired by GTA/PRoPE with sparse routing inspired by FasterVGGT to condition cross-view interactions on camera geometry. A zero-initialized output projection preserves the pretrained model’s behavior at initialization, while the auxiliary design leaves the pretrained attention and positional representation unchanged and keeps additional computational cost low.

Critical strengths

Main strengths supporting acceptance:

  • efficient and practically relevant strategy for exploiting known camera calibration, supported by controlled comparisons, informative ablations, and improvements across several recent reconstruction backbones.
  • the complementary calibrated-geometry evaluation and experiments on autonomous driving and real-world robotic manipulation further strengthen the paper.

Detailed list of strengths:

  1. The adaptation strategy is interesting and efficient. The authors provide provides an informative, controlled comparison with other camera-conditioning strategies, with CalibAdapter outperforming the alternative strategies on different metrics and most datasets (except ETH3D on which none of the strategy seems to work).

  2. The proposed Calibrated Pointmap Error (CPE) aiming to evaluate depth errors as metric 3D discrepancies under the supplied GT calibration is an interesting complementary metric.

  3. The authors show that the proposed strategy can be applied with success to several recent feedforward models such as VGGT-Ω, VGGT, DA3-Small, and π³,

  4. The authors provide design choice analyses in Section 4.4 which show primarily that projective conditioning significantly improves performance over a parameter-matched plain adapter, whereas sparse routing provides only small gains. Furthermore, early-to-middle insertion gives the best CPE and F-score among the tested positions, and adding additional branches can further improve accuracy but it comes at the cost of increased number of parameters to be trained. According to moderate sparsification preserves accuracy while substantially reducing the additional FLOPs of dense projective interaction, which also supports the lightweight design.

  5. Finally, I appreciated the fact that CalibAdapt and the alternative calibration conditionings were also evaluated on nuScenes following DriveVGGT’s data-processing and evaluation protocol and on a real-world robot manipulation setting.

Critical weaknesses

Main concerns against acceptance are:

  • incremental architectural novelty and incomplete adaptation baselines
  • insufficient implementation details and the absence of comparisons with LoRA and backbone fine-tuning leave the source of CalibAdapter’s advantage partly unresolved.

Weaknesses in more details:

  1. The architectural novelty is somewhat limited. Although the adaptation strategy is interesting and efficient, it mainly combines projective camera transformations inspired by GTA/PRoPE, sparse routing inspired by FasterVGGT, and an established zero-initialized residual adaptation mechanism. These components are integrated into an auxiliary branch that leaves the pretrained attention and positional representation unchanged. The contribution therefore lies primarily in their integration for calibration-aware adaptation of feed-forward reconstruction models.

  2. I really like the fact that Tables provide controlled comparison of different camera-conditioning strategies (+ MapAnything-style, + DA3-style, + PromptDA-style) . Using the same frozen VGGT-Ω backbone and matched trainable-parameter budgets allows comparing different manners of incorporating calibration. Indeed, the best results are obtained with the proposed CalibAdapter. However, there are several open questions concerning these experiments. How the 5.248M parameters are allocated in each alternative modules? The authors should add architectural details allowing to better illustrate how those models were implemented. Different “styles” might require different amount of “ parameter budget” to be optimal. The frozen-backbone comparison is informative, but it remains unclear whether this setting favors the alternative styles, hence maybe adding an experiment with backbone fine-tuning would allow clarifying this.

  3. The paper somewhat lacks comparison against standard alternatives such as LoRA with the same adaptation budget. Such a comparison would further strengthen the paper. Same, it would be necessary to compare with a full fine-tuning of different styles including the proposed one.

  4. The real-world robotic evaluation while interesting is rather small and limited, covering only three tasks with 20 trials per task.

Questions for authors

In Table 1(b) the authors attach CalibAdapter to VGGT, π³, and DA3-Small, and compare each adapted model with its original version, where a separate adapter is trained for each backbone on ScanNet-v2. While this experiment demonstrates that the adaptation approach applies across different architectures; it does not demonstrate that a single trained adapter transfers between backbones without retraining.

The author's should cite Wonbong et al, Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors, CVPR'25 as being amongst the first paper proposing to integrate depth, camera priors in a feed-forward 3DFM models.

7 comments
AuthorsOpenReview ↗

Response to Reviewer iKUh — W1: Methodological Novelty and Practical Contribution

We thank the reviewer for the constructive feedback and for recognizing the practical relevance, efficiency, and extensive experimental validation of our work. We address your concerns point by point below.

W1: Methodological Novelty and Practical Contribution

We agree that projective attention (GTA/PRoPE), sparse routing (FasterVGGT), and zero-initialized residual adaptation are established techniques. As the reviewer noted, our contribution lies in their targeted integration for calibration-aware adaptation, motivated by a practical problem.

In real-world applications such as robotic manipulation and autonomous driving, camera intrinsics and extrinsics are often known. Nevertheless, pretrained 3D foundation models may still produce substantial metric geometric errors under the provided calibration. Our work focuses on efficiently adapting these models to known-camera settings without costly backbone retraining.

To address this problem, CalibAdapter introduces explicit projective geometry into cross-view feature interactions through a lightweight residual branch, preserving the original pretrained attention pathway. Unlike input-level, camera-token, or decoder-level conditioning, our design directly incorporates calibrated geometry into cross-view feature interactions under the same trainable-parameter budget.

Our controlled ablations (Table 3(a)) support this design choice. Moreover, the extended cross-backbone evaluation (Table 1(b)) demonstrates that CalibAdapter further improves official MapAnything, DA3-Small, and Pi3X, even when their native camera-conditioning mechanisms already receive GT K/TK/T.

Together, these results highlight the practical relevance of calibration-aware adaptation and the effectiveness of our targeted methodological design.

AuthorsOpenReview ↗

Response to Reviewer iKUh — W2: Architectural Details and Parameter Allocation

W2: Architectural Details of Camera-conditioning Strategies

To clarify the implementation of the four camera-conditioning strategies in Table 1(a), we provide their architectural configurations and parameter allocations below. All four strategies use the same pretrained VGGT-Ω\Omega backbone with frozen original weights and prediction heads, and an identical budget of 5.248M trainable parameters.

(1) MapAnything-style. The input-level conditioning pathway is inspired by MapAnything. GT intrinsics are converted into camera-frame ray maps and encoded by a Ray CNN, while relative rotations (quaternions) and normalized translations are processed by separate MLPs. These camera features are fused with VGGT-Ω\Omega image patch features through addition and LayerNorm before the original attention blocks.

The 5.248M trainable parameters comprise the Ray Encoder (3.926M), Rotation Encoder (0.660M), Translation Encoder (0.660M), and Fusion LayerNorm (0.002M).

(2) DA3-style. The camera-token conditioning pathway is inspired by DA3, encoding normalized camera poses and intrinsics into 9D representations (translation, quaternion, and two FoVs). A lightweight Camera Encoder (width 320, four Transformer blocks) processes these features, followed by a 320→192→1024320\rightarrow192\rightarrow1024 projection bridge. The resulting camera tokens replace the corresponding token slots before block B4 of VGGT-Ω\Omega.

The 5.248M trainable parameters comprise the Camera Encoder (4.989M) and projection bridge (0.259M).

(3) PromptDA-style. Decoder-level conditioning is inspired by PromptDA. GT K/TK/T are encoded into 19D spatial camera maps (3D rays and a 16D pose representation), which are processed by four independent three-layer CNN branches. The resulting camera features are added to the corresponding DPT refinement features at four decoder stages, after the native residual convolutions and before upsampling. The final convolution in each branch is zero-initialized.

The 5.248M trainable parameters are distributed across RefineNet4 (1.549M), RefineNet3 (1.466M), RefineNet2 (1.264M), and RefineNet1 (0.970M).

(4) CalibAdapter (Ours). We introduce a parallel camera-conditioned attention branch at the B7 global attention block of VGGT-Ω\Omega. Patch features are projected into Q/K/V (16 heads, 64 dimensions each), transformed using projective operators derived from GT K/TK/T, and processed by sparse cross-view attention. Pooled routing (128-query/64-key token blocks) selects the top 25% key blocks. The attention outputs are transformed back and added as a residual without replacing the original attention.

The 5.248M trainable parameters comprise the Input Linear (1.050M), Q/K/V projections (3.149M), Output Linear (1.050M), and Q/K normalization (256 parameters). The input projection is identity-initialized, Q/K/V and Q/K normalization are initialized from the pretrained B7 weights, and the output projection is zero-initialized.

Regarding Parameter-budget Fairness. We agree that different conditioning strategies may benefit from different parameter budgets. Our unified 5.248M budget is designed to enable a controlled comparison under the same parameter-efficient adaptation constraint. To further examine the influence of the frozen-backbone setting, we additionally conduct full fine-tuning experiments for all four strategies, as reported below.

AuthorsOpenReview ↗

Response to Reviewer iKUh — W3: LoRA and Full Fine-tuning Comparisons

W3: LoRA and Full Fine-tuning Comparisons

We conduct additional comparisons with parameter-matched LoRA and full fine-tuning to further assess the effectiveness of CalibAdapter.

(1) Parameter-matched LoRA Baseline

To assess whether the improvements can be achieved through standard parameter-efficient fine-tuning, we additionally evaluate LoRA on the same VGGT-Ω\Omega backbone under the ScanNetV2 10K-step training protocol.

Specifically, LoRA is applied to the QKV, attention output, and FFN projections across 10 global blocks, with rank r=32r=32, scaling factor α=32\alpha=32, and zero dropout. This yields 5.243M trainable parameters, closely matching CalibAdapter's 5.248M budget, while keeping the pretrained backbone and prediction heads frozen.

MethodCPE ↓AbsRel ↓RMSE ↓Acc. ↓Comp. ↓CD ↓F-score ↑
VGGT-Ω\Omega (Frozen)0.22400.10540.28770.17900.11940.29840.4563
+ LoRA0.21000.09520.24570.14130.11920.26050.4968
+ CalibAdapter (Ours)0.16680.08120.20130.10140.12010.22150.4730

This comparison examines whether explicit camera-conditioned adaptation provides additional benefits beyond generic parameter-efficient fine-tuning.

(2) Full Fine-tuning Comparison

To examine whether the frozen-backbone setting favors CalibAdapter, we additionally conduct full fine-tuning of all four camera-conditioning strategies on ScanNetV2. All adapted methods use the same pretrained VGGT-Ω\Omega initialization, 10K training steps, and strict312 evaluation protocol. The backbone and corresponding conditioning modules are jointly optimized.

Method (Full FT, 10K)CPE ↓AbsRel ↓RMSE ↓Acc. ↓Comp. ↓CD ↓F-score ↑
VGGT-Ω\Omega (Frozen)0.22400.10540.28770.17900.11940.29840.4563
+ MapAnything-style0.15280.07470.18810.09960.10290.20250.5073
+ DA3-style0.13890.06480.16880.08580.09090.17670.5528
+ PromptDA-style0.15550.07990.18900.10120.10400.20520.5050
+ CalibAdapter (Ours)0.12970.06590.16500.08160.09330.17490.5382

CalibAdapter achieves the lowest CPE (0.1297), reducing CPE by 6.64% relative to the second-best DA3-style (0.1389). It also achieves the best RMSE, point-cloud accuracy, and Chamfer Distance, while DA3-style performs better on AbsRel, completeness, and F-score.

These results demonstrate that CalibAdapter's advantage in calibrated geometric reconstruction persists under full backbone fine-tuning, rather than arising solely from the frozen-backbone setting. Our primary contribution remains the effectiveness of explicit projective conditioning for parameter-efficient adaptation.

Reviewer_JG2jOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a way to adapt 3D foundation models like VGGT (Omega), Pi3, DA3 etc. to scenarios with known camera intrinsics and extrinsics. To achieve this, the authors introduce a new module CalibAdapter, which adds a zero-initialized residual attention branch to the frozen model as a lightweight adapter to inject the calibration information. To validate the calibrated setting, the authors introduce CPE, a metric 3D error to measure the agreement of the predicted depth to the GT cameras and poses. Extensive experiments and ablations validate the effectiveness and design choices of the proposed method.

Critical strengths

Practical and clear solution: Some of the recently released 3D foundation models do not provide native support for calibrated camera or known pose settings. This paper provides a clear formulation for how to adapt such models with minimal training to such settings with extensive ablations for parameter matched alternatives and design choices. The method seems to be effective in aligning the model with prior calibration and poses given at inference.

Critical weaknesses

Training free baseline: From my understanding, the CPE measures two things: (1) the scale consistency between the predicted depth and predicted poses and (2) the residual per pixel depth error. From table 10 it seems that scale (1) is the most important factor. I wonder if there is a training free baseline that uses the GT calibration and poses to recover the scale could give most, if not all the improvements proposed in the paper. One example could be to scale the predicted depth such that the scale aligns to gt poses using matches or reprojection errors. I would like to see at least one such baseline in the paper tables. In addition, Table 10 shows in domain results only, showing oracle-scale results would clarify domain adaption to genuine use of the calibrated inputs.

Questions for authors
  • Training free baseline: Could you provide an training free baseline based on reprojection errors or matches to test the effectiveness of the fine-tuned method?

  • Out of domain oracle scale: Do the improvements in table 10 come mainly from the models adaption to the domain or actual use of the calibrated inputs?

  • Single intrinsics datasets: All training and evaluation datasets are single intrinsics and relatively similar (scannet, ARKit both iPhone cameras). How does the fine-tuning effect results on datasets with in the wild cameras like MegaDepth?

  • Pose accuracy and instrincis: This is more a clarification question. How well do the predicted poses and intrinsics align with the gt after the fine-tuning?

6 comments
AuthorsOpenReview ↗

Response to Reviewer JG2j — W1 & Q1: Training-free Scale Calibration

We thank the reviewer for the constructive feedback and for recognizing the practical value of our method and the merits of our controlled experiments. We address your concerns below with additional experiments and clarifications.

W1 & Q1: Training-free Scale Calibration

We agree that a training-free geometric baseline is important for distinguishing improvements from scale recovery and learned adaptation.

Our original evaluation already employs training-free trajectory-based scale recovery, estimating a single global scale from predicted and GT camera trajectories without updating model parameters.

We additionally evaluate a matching-based geometric scale recovery baseline. Specifically, we first extract sparse cross-view correspondences using LightGlue and filter geometrically inconsistent matches using the known camera calibration (GT K/TK/T). We then triangulate the verified correspondences to obtain sparse metric depth anchors and robustly estimate a single global scale per multi-view tuple by aligning these anchors with the model-predicted depths. The estimated scale is applied uniformly to all predicted depths, without GT depth, per-view alignment, or additional model training.

We evaluate both Frozen VGGT-Ω\Omega and CalibAdapter under identical scale-recovery protocols on ScanNetV2 strict312:

MethodScale RecoveryCPE Mean ↓CPE P90 ↓AbsRel ↓RMSE ↓MAE ↓δ1\delta_1 ↑
VGGT-Ω\Omega (Frozen)Trajectory0.22400.46280.10540.28770.211291.28
+ CalibAdapter (Ours)Trajectory0.16680.37780.08120.20130.158091.95
VGGT-Ω\Omega (Frozen)Matching0.10100.19890.05020.17770.095696.76
+ CalibAdapter (Ours)Matching0.08620.18460.04120.13260.081698.01

We acknowledge that matching-based calibration provides a strong training-free alternative: Frozen + Matching (0.1010) outperforms CalibAdapter with our original trajectory-based recovery (0.1668).

Nevertheless, under the same matching-based recovery, CalibAdapter further reduces CPE from 0.1010 to 0.0862 (14.7%), while also improving AbsRel and RMSE. This improvement also persists under oracle global scaling (0.0850 → 0.0544), indicating improvements in residual geometry beyond scale estimation.

Together, these results demonstrate that training-free geometric scale recovery and learned camera-conditioned adaptation offer complementary benefits.

AuthorsOpenReview ↗

Response to Reviewer JG2j — Q2: Out-of-domain Oracle Scale and Calibration Dependency

Q2: Out-of-domain Oracle Scale and Calibration Dependency

We extend the oracle-scale analysis in Appendix Table 13 to all five datasets in Table 2, using the same FiveMix-20K checkpoint without target-domain fine-tuning.

MethodScale Source7ScenesDTUScanNet++ScanNetV2†ETH3D
VGGT-Ω\Omega (Frozen)Predicted Trajectory0.13220.01400.10440.22400.1511
+ CalibAdapter (Ours)Predicted Trajectory0.09640.00910.08560.18070.1608
VGGT-Ω\Omega (Frozen)Oracle Global0.07870.00840.11000.08500.1273
+ CalibAdapter (Ours)Oracle Global0.07490.00730.09250.06220.1531

† In-domain dataset. Oracle uses a single GT-depth-derived global scale per tuple.

Under oracle scaling, CalibAdapter achieves lower CPE on three of the four held-out datasets, indicating that its residual geometric improvements extend beyond in-domain evaluation and are not solely attributable to global scale recovery.

To further examine whether CalibAdapter effectively utilizes the provided camera calibration, we report the camera-input intervention from Appendix Table 14:

Camera InputCPE Mean ↓CPE P90 ↓
Correct (K/TK/T)0.16680.3778
Noisy (K/TK/T)0.17300.4017
Identity0.51071.0445
Shuffled (K/TK/T)0.70871.6290

Only the camera parameters supplied to CalibAdapter are perturbed; the original GT K/TK/T are retained for scale recovery and evaluation.

Mild calibration noise causes limited degradation, whereas shuffling the camera inputs increases CPE from 0.1668 to 0.7087. Since the model weights and evaluation protocol remain unchanged, this provides direct evidence that CalibAdapter's predictions depend on the supplied camera calibration, rather than ignoring it during adaptation.

Together, these experiments demonstrate residual geometric improvements across multiple held-out domains and confirm the model's dependence on calibrated camera inputs.

AuthorsOpenReview ↗

Response to Reviewer JG2j — Q3: Generalization to In-the-wild Cameras

Q3: Generalization to In-the-wild Cameras

To evaluate whether CalibAdapter's improvements generalize to in-the-wild cameras, we conduct a zero-shot evaluation on MegaDepth-1500 using the FiveMix-20K checkpoint, without any target-domain fine-tuning.

MethodNormalized CPE ↓AbsRel ↓Normalized RMSE ↓
VGGT-Ω\Omega (Frozen)0.15340.07840.2126
+ CalibAdapter (Ours)0.14120.07360.1971

This evaluation covers 1,500 image pairs from two scenes. As MegaDepth's reference cameras and depths are derived from SfM/MVS, we report scene-baseline-normalized CPE, rather than metric CPE in meters.

CalibAdapter also improves AbsRel and normalized RMSE by 6.13% and 7.29%, respectively. The results provide additional evidence of zero-shot generalization to in-the-wild imagery.

Reviewer_zzkhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors look at what happens when depth from feed-forward 3D models (VGGT, π3, DA3) is back-projected with the known camera calibration that driving and robotics systems already have. To measure the error, they define CPE: one scale per sequence is recovered from the predicted vs. given camera trajectory, the scaled depth is back-projected with the given K and T, and the per-pixel 3D error is averaged. To reduce CPE, they freeze the backbone and add one small residual attention branch next to an early cross-view block. In this branch, Q/K/V are transformed with the relative camera geometry (PRoPE/GTA style), and pooled top-k routing (from FasterVGGT) keeps the cost down. CPE drops on four backbones and on most held-out datasets. They also report gains on nuScenes, and on a dual-arm robot success goes from 68.3% to 86.7% against a plain adapter with the same parameter count.

Critical strengths
  1. This work addresses a practical issue. Most deployed systems have calibrated cameras for free, and a cheap way to make frozen 3D foundation models use them is useful. The adapter is small, leaves the backbone untouched, and works on four different backbones.
  2. The internal controls are sensible. The parameter-matched plain adapter, the oracle-scale diagnostic (Table 10), and the shuffled/identity camera inputs (Table 11) show that the gain over the backbone comes from the calibration itself, and not just from extra capacity or a better scale estimate.
Critical weaknesses
  1. The authors claim that existing calibration-aware methods do not guarantee that predicted depth back-projects to geometry consistent with the given cameras. However, I did not find experiments that verify this. The “-style” baselines are the authors’ re-implementations on a frozen VGGT-Ω, and VGGT and π3 receive no camera input. The only released model presented as camera-aware is DA3-Small, and the paper does not say whether it was given the ground-truth poses at test time. Open-source models that accept known cameras at inference are available, e.g. MapAnything [1], DA3 [2], and Pi3X [3], the camera-conditioned version of π3, which is one of the paper’s own backbones. Their CPE under the same calibration should be reported, and CalibAdapter should be compared against them. [1]: https://arxiv.org/abs/2509.13414 [2]: https://arxiv.org/abs/2511.10647 [3]: https://github.com/yyfz/Pi3

  2. I don’t think there is much new here. CPE turns out to be a depth error after one scale fit (the extrinsics cancel, see Appendix A.1), and most of the frozen model’s CPE is just a wrong scale (0.224 vs. 0.085 with the oracle scale, Table 10). The adapter combines existing components: projective attention from GTA/PRoPE, a zero-init residual branch, and block-sparse routing from FasterVGGT.

Questions for authors
  1. What CPE do MapAnything [1], DA3 with pose conditioning [2], and Pi3X [3] get when given the same ground-truth calibration, and how does CalibAdapter compare with them?
  2. Was DA3-Small in Table 1(b) given the ground-truth poses at test time?
  3. The main results use seed 42, the best of the three seeds in Table 8. Could you report the mean over seeds?

I would consider raising my rating if the authors can address my above concerns.

7 comments
AuthorsOpenReview ↗

Response to Reviewer zzkh — W1 & Q1: Official Camera-conditioned Models

We thank the reviewer for the constructive feedback and for recognizing the practical value of our work and the merits of our controlled experiments. We address your concerns point by point below, supported by additional experiments and clarifications.

W1 & Q1: Comparison with Official Camera-conditioned Models

We thank the reviewer for highlighting the importance of evaluating official camera-conditioned models.

We additionally evaluate the official pretrained MapAnything, DA3-Small, and Pi3X on ScanNetV2. All three models receive GT K/TK/T through their native camera-conditioning interfaces during inference, both with and without CalibAdapter. Notably, DA3-Small in our original Table 1(b) already used GT K/TK/T during inference. For the adapted variants, we retain the pretrained backbones and native camera-conditioning mechanisms, while training CalibAdapter separately for each model.

MethodCPE Mean ↓CPE P90 ↓AbsRel ↓RMSE ↓Acc. ↓Comp. ↓F-score ↑
MapAnything0.23350.55840.11430.28570.17370.15820.3436
MapAnything+Ours0.10490.22500.05390.15170.08090.07460.5725
DA3-Small0.21160.51500.10090.26080.13760.16490.4141
DA3-Small+Ours0.15740.35500.07670.20540.10080.12810.4603
Pi3X0.08800.18060.04500.11850.05470.06600.6089
Pi3X+Ours0.07230.14470.03640.10590.04630.05470.6840
VGGT-Ω\Omega (Frozen)0.22400.46280.10540.28770.17900.11940.4563
VGGT-Ω\Omega+Ours0.16680.37780.08120.20130.10140.12010.4730

These additional evaluations demonstrate that CalibAdapter provides complementary geometric improvements to existing camera-aware foundation models. Meanwhile, Table 1(a) provides a controlled comparison of different camera-conditioning strategies under the same pretrained VGGT-Ω\Omega backbone, training protocol, and parameter budget.

These results and the corresponding evaluation protocols have been incorporated into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer zzkh — Q2: GT Camera Conditioning in DA3-Small

Q2: GT Camera Conditioning in DA3-Small

We confirm that DA3-Small in Table 1(b) receives GT camera intrinsics and extrinsics through its native camera-conditioning mechanism during inference.

CalibAdapter is introduced on top of the existing camera conditioning, rather than replacing it, further reducing CPE from 0.2116 to 0.1574 (25.6%). This demonstrates that our projective cross-view adaptation provides complementary geometric improvements even when GT camera calibration is already incorporated. We have clarified this inference configuration in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer zzkh — W2(a): CPE Interpretation and Practical Significance

W2(a): Interpretation and Practical Significance of CPE

We agree that, as derived in Appendix A.1, CPE is mathematically equivalent to ray-weighted depth error under shared calibrated cameras. Its contribution lies in a strict evaluation protocol for pixel-corresponded metric 3D accuracy, rather than a new distance function.

Unlike nearest-neighbor-based point-cloud metrics such as Chamfer Distance and F-score, CPE preserves pixel-to-point correspondences. Our evaluation protocol permits only a single global scale per multi-view tuple, without per-image scale-shift fitting or post-hoc Sim(3)/ICP alignment. This protocol is particularly relevant to calibrated robotic manipulation and autonomous driving applications, where accurate metric 3D localization is essential for downstream perception and decision-making.

Regarding scale estimation, we acknowledge its substantial influence: oracle scaling reduces Frozen VGGT-Ω\Omega's CPE from 0.2240 to 0.0850. However, under the same oracle-scale fitting protocol, CalibAdapter further reduces CPE from 0.0850 to 0.0544 (36.0%), demonstrating improvements in residual depth geometry beyond global scale correction. Importantly, oracle scaling uses GT depth and serves solely as a diagnostic reference; it is not available in practical deployment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.