- 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:
-
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).
-
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.
-
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,
-
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.
-
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:
-
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.
-
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.
-
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.
-
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
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- 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- 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 projection bridge. The resulting camera tokens replace the corresponding token slots before block B4 of VGGT-.
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 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-. Patch features are projected into Q/K/V (16 heads, 64 dimensions each), transformed using projective operators derived from GT , 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.
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- 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 , scaling factor , 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.
| Method | CPE ↓ | AbsRel ↓ | RMSE ↓ | Acc. ↓ | Comp. ↓ | CD ↓ | F-score ↑ |
|---|---|---|---|---|---|---|---|
| VGGT- (Frozen) | 0.2240 | 0.1054 | 0.2877 | 0.1790 | 0.1194 | 0.2984 | 0.4563 |
| + LoRA | 0.2100 | 0.0952 | 0.2457 | 0.1413 | 0.1192 | 0.2605 | 0.4968 |
| + CalibAdapter (Ours) | 0.1668 | 0.0812 | 0.2013 | 0.1014 | 0.1201 | 0.2215 | 0.4730 |
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- 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- (Frozen) | 0.2240 | 0.1054 | 0.2877 | 0.1790 | 0.1194 | 0.2984 | 0.4563 |
| + MapAnything-style | 0.1528 | 0.0747 | 0.1881 | 0.0996 | 0.1029 | 0.2025 | 0.5073 |
| + DA3-style | 0.1389 | 0.0648 | 0.1688 | 0.0858 | 0.0909 | 0.1767 | 0.5528 |
| + PromptDA-style | 0.1555 | 0.0799 | 0.1890 | 0.1012 | 0.1040 | 0.2052 | 0.5050 |
| + CalibAdapter (Ours) | 0.1297 | 0.0659 | 0.1650 | 0.0816 | 0.0933 | 0.1749 | 0.5382 |
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.
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/T.
Together, these results highlight the practical relevance of calibration-aware adaptation and the effectiveness of our targeted methodological design.