acceptodds
Under review as a conference paper at ICLR 2027

Boosting Metric Depth Completion via Training-Free Adaptive Response Geometry

Abstract

Depth completion aims to recover dense metric depth from sparse sensor measurements, increasingly leveraging visual foundation models as geometric priors. However, aligning these priors to true metric scale typically relies on rigid affine assumptions in predefined coordinate systems, leaving systematic calibration errors. Linearity in depth calibration depends on the response coordinate. We introduce adaptive response geometry, which makes the fixed choice of depth, log depth, or disparity an image-level unknown. A continuous response family unifies these coordinates and defines an explicit depth-dependent gain. We derive the response-gradient relation and estimate the response parameters in metric space. Hard-Dirichlet residual reconstruction completes the calibrated prior. Under deliberately incomplete metric observations, the training-free pipeline achieves macro AbsRel and macro NMed , improving both aggregate measures over PriorDA, LDCM, and Any2Full. Linearity diagnostics examine how the selected response changes the depth relation and its metric error.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
monocular depth estimation, sparse depth completion, metric depth estimation, 3D geometry, geometric fidelity, depth calibration, response geometry

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_MJu5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a training-free depth completion method that adapts the response coordinate for global metric calibration. It combines Box–Cox alignment with Hard-Dirichlet residual reconstruction to recover dense depth from a monocular prior and sparse observations.

Critical strengths

The formulation is compact and offers a well-motivated approach to depth-dependent calibration. The experiments cover multiple datasets and monocular priors, with stage ablations and an explicit discussion of the depth–normal trade-off.

Critical weaknesses

Adaptive calibration minimizes metric-depth squared error, whereas GLog and GDisp fit in their respective coordinates, confounding response flexibility with residual weighting (Eq. 17; Appendix A.4). Despite the reported gains across 45 prior–subset combinations, fixed-λ controls using the same metric objective and crossed ablations with an identical reconstruction backend are missing; the Poisson baseline also differs in constraints and solving resolution.

Questions for authors
  1. How much benefit remains when λ is fixed to −1, 0, or 1 using the same metric objective and feasibility constraints, and all calibration variants share the same reconstruction backend?

  2. Can you provide per-observation-mode results and anchor counts, clarify common valid samples, pixels, and failure handling, and quantify paired performance uncertainty with scene- or video-grouped analysis?

  3. What evidence favors hard over soft or robust constraints under identical guidance, particularly for preserving normal quality and handling point-sampled, noisy, boundary-mixed, or misregistered observations?

  4. How do you handle equivalent or weakly identifiable λ values, initialization sensitivity, and error amplification near the domain boundary, especially outside the observed depth range?

2 comments
AuthorsOpenReview ↗

Official Comment by Authors for Reviewer MJu5

Dear Reviewer MJu5,

Thank you for the rigorous review and for highlighting our compact depth-dependent calibration formulation, broad evaluation across priors and datasets, and explicit analysis of the depth–normal trade-off. We have completed intensive experiments addressing the raised concern. Due to the limited computational resources and strict timeline, we report the depth completion performance metrics on a randomly chosen Barn dataset. We will also update the full results on a wider range of datasets in the revised version of the main paper.


Q1: Clarification on Fixed-Controls and Crossed Backend Ablations Thank you for suggesting a controllable experiments on different baseline involving representative λ\lambda. We report these results in the table below.

Calibration Variant (Barn, all minimizing Eq. 17)AbsRel+ Hard-Dirichlet AbsRel
Fixed λ=−1\lambda = -1 (disparity)0.08240.0480
Fixed λ=0\lambda = 0 (log-depth)0.10410.0559
Fixed λ=+1\lambda = +1 (metric depth)0.12350.0625
Adaptive λ\lambda (Ours)0.03010.0251

Under identical fitting objectives, feasibility constraints, and refinement backends, adapting λ\lambda reduces AbsRel by 63.4%–75.6% before completion and 47.7%–59.8% after Hard-Dirichlet completion. This indeed reveals the strong motivation of our proposed method. We will update the manuscript to include more discussion and related experimental results.


Q2: Per-Mode Results, Anchor Counts, Valid Pixels, and Failure Handling

We initially detailed this information on different parts in both the main paper and appendix (A.3, F, G.3, and H). Here we provide a more complete summary as below.

  • Samples: 3,698 images and 11,094 image–mode entries across 9 subsets.

  • Anchor counts: Median valid anchor counts per subset across the three modes (Distributed / Central removal / Outer removal) are reported in Table H (e.g., 4,207 / 2,275 / 1,941 on Barn).

  • Per-mode results (9-subset Macro p50 AbsRel): Ours achieves 0.0212 / 0.0344 / 0.0433; PriorDA achieves 0.0188 / 0.0353 / 0.0683; LDCM achieves 0.0163 / 0.0480 / 0.0731; Any2Full achieves 0.0190 / 0.0415 / 0.0640. Our gains are most pronounced under structured spatial missingness (Central and Outer removal).

  • Common valid pixels and failure handling: All methods are evaluated over the exact same pixel set (valid reference pixels inside the scoring mask, excluding anchor pixels). Ours records 0 failures across all 45 prior.


Q3: Hard vs. Soft/Robust Constraints under Identical Settings

Under identical Stage-1 parameters and gradient guidance, we replace exact Dirichlet elimination with a soft anchor penalty μobs∑(xp−log⁡sp)2\mu_{\mathrm{obs}} \sum (x_p - \log s_p)^2, and we report below the depth completion performance on Barn dataset.

Anchor Constraint Formulationp50 AbsRelp50 NMed
Ours w/o Dirichlet (μobs=0\mu_{\mathrm{obs}} = 0)0.03019.07°
Soft constraint (μobs=0.01\mu_{\mathrm{obs}} = 0.01)0.02959.11°
Soft constraint (μobs=0.1\mu_{\mathrm{obs}} = 0.1)0.02799.86°
Soft constraint (μobs=1.0\mu_{\mathrm{obs}} = 1.0)0.025810.63°
Hard Dirichlet (exact elimination)0.025110.98°

From these results one can summarize as

  • Depth–normal trade-off: As the anchor penalty softens, metric depth error degrades monotonically toward Ours w/o Dirichlet, whereas surface-normal error (NMed) moves smoothly in the opposite direction—directly illustrating the depth–normal trade-off analyzed in the main text.
  • Regularization sensitivity: Sensitivity to (λdata,λlap)(\lambda_{\mathrm{data}}, \lambda_{\mathrm{lap}}) is reported in Appendix D (Table D.1) and remains stable across {10−4,10−3,10−2}2\{10^{-4}, 10^{-3}, 10^{-2}\}^2.

We also want to enhance the conclusion by mentioning the experiments suggested by Reviewer rKr7 with lidar-scan-like noise. Under heavy measurement noise (4 cm/12 cm), replacing hard elimination with a soft constraint (μobs=0.5\mu_{\mathrm{obs}} = 0.5) lowers p50 AbsRel from 0.0178 to 0.0147.

We will revise the main paper to report both findings in the noise-rich and noise-sparse cases for hard/soft constraints.


message follows.

AuthorsOpenReview ↗

Official Comment by Authors for Reviewer MJu5 (2)

Q4. Weakly Identifiable λ\lambda, Initialization Sensitivity, and Domain Boundaries

We thank the reviewer for the insightful question. We have done a few works to stablize the computation and avoid several degenerated cases.

  • Scale centering and numerically smooth transition (Eqs. 5–6, Appendix A.1): Normalizing by the anchor geometric mean s0s_0 enforces ϕλ(s0)=0\phi_\lambda(s_0) = 0 and ϕλ′(s0)=1/s0\phi_\lambda'(s_0) = 1/s_0 for every λ\lambda, decoupling λ\lambda from overall global scale. Evaluating the power branch via expm1 and smoothly transitioning to the logarithmic branch when ∣λ∣|\lambda| is tiny eliminates catastrophic cancellation near λ→0\lambda \to 0.

  • Weak identifiability and parameter drift (Eq. 21): When the observed depth span is narrow or the underlying relationship is nearly linear, neighbouring (λ,α,β)(\lambda, \alpha, \beta) tuples induce virtually indistinguishable monotonic curves over the observed interval. We address this weak identifiability via damped 3×33 \times 3 Levenberg–Marquardt optimization, accepting a step only when it strictly decreases the metric residual to suppress parameter drift.

  • Domain boundaries and out-of-range extrapolation (Appendix A.1): Because α>0\alpha > 0 and ϕλ\phi_\lambda is strictly increasing, g(h)=1+λ(αϕλ(h)+β)g(h) = 1 + \lambda(\alpha\phi_\lambda(h) + \beta) is positive across the entire dense prior if and only if it is positive at the global extrema hmin⁡h_{\min} and hmax⁡h_{\max}. Checking positivity at these two endpoints before accepting every LM step prevents singularities and unbounded error amplification outside the observed depth range, resulting in 0 failures across all 45 combinations.

Reviewer_ppaJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a training-free method for metric depth completion. It first fits a three-parameter global mapping to sparse depth observations using the Box–Cox transformation. This allows the method to adaptively choose a response space and calibrate monocular depth predictions to metric scale. It then refines the result spatially by reconstructing log-depth residuals under hard Dirichlet constraints. The goal is to correct systematic errors that remain after affine calibration in fixed depth, log-depth, or inverse-depth spaces.

Critical strengths

Simple and interpretable design. Global calibration estimates only three parameters and does not require fine-tuning the depth model. Separating this step from local residual refinement also makes it clear what each component contributes.

Critical weaknesses
  1. The methodological novelty and advantages over closely related work remain insufficiently established.

Prior work has explored ideas closely related to both stages of this method. Yang et al. [1] use sparse 3D observations and monotonic spline fitting to calibrate monocular depth. Recovering metric depth through a monotonic nonlinear mapping is therefore not new. The Midas Touch for Metric Depth [2] also uses a two-stage approach: scale recovery followed by pixel-wise refinement. The paper further acknowledges that its reconstruction with hard Dirichlet constraints builds on OASIS-DC [3].

These similarities do not mean the methods are identical. However, they leave a key question unanswered: what advantage does the proposed three-parameter mapping offer over existing nonlinear calibration and spatial refinement methods? Using fewer parameters may provide useful regularization, but it does not necessarily improve accuracy, robustness, or extrapolation. The current analysis does not clearly demonstrate these benefits. I therefore find it difficult to identify a methodological advance beyond choosing a particular mapping family and combining existing ideas.

  1. Missing comparisons with the closest prior methods prevent a convincing assessment of the proposed contribution.

The paper cites and discusses these related methods, but does not compare directly with Yang et al.’s monotonic spline calibration [1], The Midas Touch for Metric Depth [2], or OASIS-DC [3]. The fixed-space calibration, KNN, LWLR, and Poisson baselines do not replace comparisons with these methods.

The most important comparison is between the proposed mapping and existing monotonic nonlinear mappings. These tests should use the same monocular depth prior, sparse observations, and residual refinement backend. Where applicable, they should also use the same fitting objective in metric depth space. The current evidence therefore does not establish a clear performance or robustness advantage over the most closely related approaches.

[1] Yang, Steven, et al. "Zero-shot metric depth estimation via monocular visual-inertial rescaling for autonomous aerial navigation." 2026 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2026. [2] Ma, Yu, et al. "The Midas Touch for Metric Depth." arXiv preprint arXiv:2605.11578 CVPR 2026 (2026). [3] Cho, Jaehyeon, and Jhonghyun An. "OASIS-DC: Generalizable Depth Completion via Output-level Alignment of Sparse-Integrated Monocular Pseudo Depth." arXiv preprint arXiv:2602.01268 ICRA 2026 (2026).

Questions for authors
  1. What is the key advantage of the proposed three-parameter mapping over the monotonic spline calibration in [1]? Can theoretical analysis or controlled experiments demonstrate better stability with sparse observations or better extrapolation beyond the observed depth range?

  2. Can the authors include direct comparisons with [1] and [2]? For [3], could the authors at least compare against its non-learning-based pseudo-depth construction stage to clarify the incremental contribution of the proposed spatial reconstruction?\

  3. Could the authors compare fixed-space calibration, monotonic spline calibration, and the proposed mapping under a consistent fitting objective and an identical refinement backend, while varying the number of observations, their depth-range coverage, and measurement noise?

3 comments
AuthorsOpenReview ↗

Official Comment by Authors for Reviewer ppaJ

Dear Reviewer ppaJ,

Thank you very much for the careful reading and the constructive comments. We thank you for acknowledging our simple, training-free method. Here we provide our answer to your listed concerns.


Q1. Methodological Distinctions and Advantages over [1, 2, 3]

  • Comparison with Yang et al. [1]: [1] rescales monocular predictions at visual-inertial feature points by first converting depth into a fixed disparity representation and fitting hand-selected function families (polynomials, exponentials, or cubic penalized B-splines). Several of its fixed mappings (least-squares fits in linear or logarithmic coordinates) correspond to discrete special cases of our continuous response family. Crucially, [1]'s B-spline knots are defined solely over the value interval of observed anchors, and its official implementation clips prior values outside that interval (np.clip), collapsing nearer or farther structures beyond the observed depth range to constant depth. In contrast, our 3-parameter formulation globally adapts the response coordinate, remains strictly monotonic and analytic across the entire prior domain, and yields consistent gains across five foundation priors and multiple scene types (Appendix E).
  • Comparison with The Midas Touch [2]: [2] segments the RGB image into superpixels, fits local affine parameters via ordinary least squares (OLS) inside superpixels with sufficient 3D anchors, propagates estimates across empty superpixels via a graph, and applies post-hoc smoothing. Partitioning in RGB space is ad hoc—texture boundaries need not correspond to calibration drift—while each local fit remains fixed-space OLS. We instead introduce a unified 3-parameter global framework that subsumes fixed-space least squares as special cases, achieving higher accuracy with far fewer parameters.

  • Comparison with OASIS-DC [3]: While [3]'s non-learned stage uses a related Poisson-style step mid-pipeline, its global alignment relies on fixed linear-depth affine fitting (equivalent to λ=1\lambda = 1). Our Stage 1 is orthogonal to [3]: replacing [3]'s linear alignment with our 3-parameter adaptive calibration directly improves its pseudo-depth (see table below). The two designs are complementary rather than conflicting.

From our side, our main contribution is a three-parameter global calibration that learns the depth coordinate in which to align a monocular prior to sparse metric anchors. It unifies disparity, log-depth, and metric-depth affine fitting as special cases and preserves depth ordering across the full prior range, including beyond the observed anchors. On Barn, adaptive calibration reduces AbsRel from 0.0824 for the best fixed-coordinate variant to 0.0301 before any completion backend. Replacing OASIS-DC’s linear alignment with our calibration also improves its pseudo-depth, demonstrating that the front end can benefit an existing completion pipeline (Table for Reviewer rKr7).


message follows.

AuthorsOpenReview ↗

Official Comment by Authors for Reviewer ppaJ (2)

Q2: Complementary experimental results.

We strictly follow your suggested evaluation criteria: identical MoGe-3 prior, identical sparse observations, identical Hard-Dirichlet backend (Eq. 23), and—where applicable—the identical metric squared-error objective (Eq. 17). We run the official implementations of all three works; because [3] does not release pretrained network weights, we compare directly against its official non-learned pseudo-depth construction stage as suggested.

Method (Barn, main protocol)Refinement stepp50 AbsRel
Yang et al. [1] (official monotonic P-spline)None0.0330
Ours w/o Dirichlet (3-param Box–Cox)None0.0301
Yang et al. [1] (disparity spline)Identical Hard-Dirichlet0.0263
Yang et al. [1] (metric-L2L_2 objective spline)Identical Hard-Dirichlet0.0265
The Midas Touch [2] (official two-stage)Official0.0448
OASIS-DC [3] (official non-learned pseudo-depth)Official linear Poisson0.0344
OASIS-DC [3] + Our Stage-1 calibrationOfficial linear Poisson0.0254
Ours (full)Identical Hard-Dirichlet0.0251

From this table, one can summarize: Without post-refinement (stage-2) our 3-parameter mapping outperforms [1]'s 14-parameter monotonic spline (0.0301 vs. 0.0330), and remains superior under the exact same metric-L2L_2 objective and Hard-Dirichlet backend (0.0251 vs. 0.0263–0.0265).

Compared with non-learned methods, our proposed method reduces p50 AbsRel by 44.0% relative to [2] (0.0448) and by 26.9% relative to [3]'s pseudo-depth stage (0.0344).

Our method is also orthogonal to [3]: simply substituting our Stage-1 calibration for [3]'s linear alignment lowers [3]'s pseudo-depth AbsRel by 26.1% (0.0344→0.02540.0344 \to 0.0254).


Q1-Q3. Sparse/Noisy Stability and Extrapolation Beyond Observed Depth Ranges

For varying observation counts and measurement noise (Q1, Q3), please see the 16/32/64-beam LiDAR sparsity sweep and ellipsoidal-covariance noise experiments in our response to Reviewer rKr7 (W1, W2): Stage 1's three global parameters remain virtually invariant across a 4×4\times anchor-density sweep and across three realistic LiDAR noise levels (0.0255–0.0259).

Regarding extrapolation and depth-range coverage, the three observation modes in our main protocol naturally probe progressively narrower coverage—specifically, Outer removal retains anchors only within the central 25% region and requires extrapolating into the outer 75%. Breaking down the controlled comparison above by mode under the same metric objective and identical Hard-Dirichlet backend:

Method (Barn, + Identical Hard-Dirichlet, p50 AbsRel)DistributedCentral RemovalOuter Removal
Fixed λ=−1\lambda = -1 (metric-L2L_2 objective)0.01880.06970.0904
Yang et al. [1] (disparity spline)0.01780.02910.0512
Yang et al. [1] (metric-L2L_2 objective spline)0.01800.03130.0546
Ours0.01790.02690.0406

When observations cover the scene well (Distributed), our 3-parameter mapping matches the 14-parameter spline (0.0179 vs. 0.0178–0.0180). As coverage narrows, the margin widens substantially: on Outer removal, Ours lowers AbsRel by 21%–26% (0.0406 vs. 0.0512–0.0546) and reduces median surface-normal angular error from 16.0° to 10.6°. This directly reflects the structural difference noted in Section 1: [1] clips prior values outside the observed anchor range, whereas our mapping remains strictly monotonic and analytic across the entire prior domain.

Thank you for the detailed response and additional experiments. The controlled comparisons make the benefits of the calibration step clearer. Replacing the OASIS-DC front end also provides useful evidence of its value. I acknowledge that the three-parameter formulation is compact and goes beyond fixed-coordinate affine fitting.  However, I remain concerned about the limited methodological novelty. The method still seems to be an incremental extension of existing nonlinear depth calibration using a classical transformation family. Overall, the response strengthens the empirical evidence but does not resolve my main concern about novelty. I will therefore keep my original score.

Reviewer_rKr7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a training-free method for depth completion that converts a dense relative-depth prior into a metric depth map using sparse or incomplete metric observations. Instead of fitting a global affine mapping in a fixed coordinate such as metric depth, log-depth, or disparity, it treats the response coordinate itself as an image-level unknown. A scaled Box-Cox transformation with one scalar parameter unifies these conventional spaces, and this parameter is jointly estimated with the global scale and offset from metric residuals using Levenberg-Marquardt optimization. The estimated global conversion is then combined with hard-constrained Poisson residual completion, where observed depths are enforced as Dirichlet constraints and the remaining residuals are propagated using log-depth gradient guidance.

Critical strengths
  1. The paper provides a clear and low-dimensional theoretical formulation that treats the response coordinate itself as an estimable image-level unknown, unifying metric depth, log-depth, and disparity as special cases.

  2. The paper offers an interesting observation that fixed-coordinate affine alignment leaves structured depth-dependent residuals and can retain noticeable metric error even when the prior is already highly linearly correlated with ground truth.

Critical weaknesses
  1. The paper evaluates sparse observations using a regular grid with central and outer removal; adding LiDAR-like, SfM-like, and other irregular patterns, together with a sweep over anchor ratios, would further test the robustness of the estimated response and affine parameters.

  2. The paper does not systematically evaluate robustness to noisy observations; although the reference depths come from real sensors and may already contain measurement and projection errors, adding controlled depth noise and outlier simulations, such as LiDAR ranging noise or SfM outliers, would further clarify the method’s behavior under realistic noisy inputs.

  3. It would be better to evaluate whether using the proposed dense output as an input or prior for PriorDA, PromptDA, and LDCM improves their performance, which would clarify whether the proposed estimation is complementary to existing learned completion pipelines.

  4. It would be better to report inference time and computational cost, including the LM estimation and Poisson completion stages, and to compare them with PriorDA, LDCM, and Any2Full.

Questions for authors

Please see weakness

3 comments
AuthorsOpenReview ↗

Official Comment by Authors for Reviewer rKr7

Dear reviewer rKr7,

We thank the reviewer for the careful reading and constructive suggestions. Below, we address your concerns on evaluating our proposed method on more challenging settings. Due to limited time and computational resources, we can only conduct all the experiment variants on a randomly chosen Barn dataset. The full results are underway on a wider range of datasets, showing a consistent tendency with Barn dataset, and we will update the main paper when finished.


Q1: Irregular LiDAR-like Patterns and Anchor-Density Sweep

We artificially created lidar-scan-like patterns by back-projecting the reference geometry into 3D and sampling a rotating 16-, 32-, and 64-beam LiDAR. This yields median valid anchor counts of ~3.4k, 7.1k, and 14.3k, constituting a sweep in anchor density. We show in the table below the main metrics quanifying the quality of the depth completion.

Barn scene — Median absolute relative error (p50 AbsRel ↓)
Bold marks the lowest error in each column. Italics mark the second-lowest.

Method16-beam32-beam64-beam
KNN0.03040.01920.0096
LWLR0.03040.03050.0313
Poisson0.02310.01700.0139
PriorDA (trained)0.01430.01120.0097
PromptDA (trained)0.97020.81530.2974
Any2Full (trained)0.01490.01100.0093
LDCM (trained)0.00990.00720.0060
Ours0.01360.00990.0070

Across all three beam densities, our proposed method, which does not rely on sophisticated training, achieves 27%–78% lower error than the three non-learned baselines (KNN, LWLR, Poisson) and consistently outperforms the learned PriorDA (by 5%–28%) and Any2Full (by 9%–24%), trailing only LDCM. As LDCM method employed a two-stage inference pipeline, wherein the second stage, a depth refinement network is trained and takes a trained-free prompt as input, our method can further enhance this procedure by replacing their stage 1. We report these results in section Q3.


Q2: Robustness to Realistic Anisotropic Noise We thank the reviewer for this important suggestion. We simulate realistic LiDAR ranging and beam divergence noise using realistic surface-normal-aligned ellipsoidal covariances. Each 3D point is perturbed by sampling from an oblate ellipsoid—thin along the surface normal (σn\sigma_{n}) and wider across the tangent plane (σt=3σn\sigma_{t} = 3 \sigma_{n}), scaling slightly with range—and then reprojected into the camera frame. This aligns with physical distortion in lidar scanning. On 32-beam LiDAR, we evaluate three noise regimes : Light (1cm / 3cm), Moderate (2 cm/6 cm), and Heavy (4 cm/12 cm).

Method (Barn 32-beam, p50 AbsRel)CleanLightModerateHeavy
KNN0.01920.01720.01840.0236
LWLR0.03050.03230.03520.0403
Poisson0.01700.01660.01770.0215
PriorDA (trained)0.01120.01090.01170.0134
PromptDA (trained)0.81530.58180.44620.4184
Any2Full (trained)0.01100.01090.01200.0168
LDCM (trained)0.00720.00780.00980.0157
Ours w/o Dirichlet (Stage 1 only)0.02590.02560.02550.0257
Ours0.00990.01040.01230.0178

Across all three noise levels, Ours outperforms all three non-learned baselines and under light noise also surpasses PriorDA and Any2Full, despite our simplification receipt and training-free nature.


message follows.

AuthorsOpenReview ↗

Official Comment by Authors for Reviewer rKr7 (2)

Q3: Serving as a Front-End for Learned Completion Pipelines

In order to test the universability of our proposed method, we select three methods with a two-stage inference pipeline, and replace each coarse front-end input with our dense output: PriorDA's global/KNN disparity condition, LDCM's coarse Poisson completion map, and PromptDA's low-resolution prompt grid. We report the depth completion metric as in the following table.

Pipeline (Barn, main protocol, p50 AbsRel)Original+ Ours Front-EndRelative Change
PriorDA0.02770.0246−11.2%
LDCM0.02130.0172−19.2%
PromptDA0.99440.0386−96.1%

Our output is complementary to learned completion pipelines: both PriorDA and LDCM improve substantially over their original versions, while PromptDA's failure under spatially missing observations is eliminated.


Q4: Inference Latency and Computational Cost

We measure completion-stage overhead given the monocular prior (excluding the prior network's forward pass) on a single NVIDIA RTX 3500 Ada GPU across 135 stratified samples spanning all nine subsets. Latency reports GPU-synchronized wall-clock time.

Method / Stage (excl. prior forward)Learned ParamsFLOPs (480×640 / 540×960)Latency (480×640 / 540×960 / 135-sample median)
Ours Stage 1: 3 × 3 LM calibration00.01 / 0.01 G2.7 / 3.5 / 3.6 ms
Ours Stage 2: Hard-Dirichlet PCG04.21 / 7.51 G11.3 / 24.8 / 22.7 ms
Ours completion total04.22 / 7.52 G14.3 / 30.5 / 26.9 ms
PriorDA completion module97.3 M414.9 / 558.7 G90.4 / 124.1 / 120.9 ms
LDCM completion module203.5 M549.3 / 1209.5 G171.4 / 451.2 / 351.3 ms

Our completion stage uses zero learned parameters, fewer FLOPs, and lower latency than the completion stages of PriorDA and LDCM.

Thank you for the additional experiments. I've carefully read the responses, and I think the rebuttals have addressed all the concerns I raised in my initial review. Given that the proposed method provides an effective complement to two-stage methods, and can refine the depth within limited computation budget, I will keep my original positive rate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.