acceptodds
Under review as a conference paper at ICLR 2027

Solving LiDAR Inverse Problems using LiDAR Diffusion Models

Abstract

Inverse problems are widely studied across imaging domains, but formulations for LiDAR under real-world corruptions remain underexplored. LiDAR is frequently corrupted by adverse weather, which we cast as a LiDAR inverse problem with explicit measurement operators. However, when 2D inverse problem frameworks are naively applied to LiDAR range views, they struggle to handle the strong spatial correlations inherent in 2.5D data. This is because they typically rely on an unrealistic assumption of pixel independence. To address this limitation, we present a new diffusion model guidance mechanism, termed correlative guidance, explicitly designed to model these missing spatial dependencies. Our framework augments the standard L2 guidance, which assumes pixel independence, with our proposed correlative term. This new guidance is implemented as a tractable, robust patch-wise Earth Mover's Distance (EMD) that compares local depth histograms. By operating on local distributions, our correlative guidance enforces local geometric consistency, a property ignored by standard L2 guidance. Extensive experiments across eight distinct corruption types demonstrate the effectiveness of our method and establish a new state-of-the-art for solving LiDAR inverse problems.

Primary area:
generative models
Keywords:
Inverse Problem, LiDAR Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_7azvOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a diffusion-based framework to solve LiDAR inverse problems i.e reconstruction of clean lidar data under real-world corruptions like adverse weather (e.g., fog, snow) and sensor flaws (e.g. missing beam, crosstalk). Diffusion posterior sampling is used to reconstruct clean data from corrupted samples. A surrogate correlative guidance using patch wise earth movers distance (EMD) is introduced along with standard L2 guidance. The authors use the method proposed in Robo3D (Kong et al) for different types of lidar degradation. The corruption operator (C) is basically a mapping from clean 3D points to corrupt 3D points for various adverse weather corruptions (missing beam, snow, fog). So the corruption is done in 3D and then the corrupted sample is projected into 2D (depth and reflectance) images which is then denoised using the proposed diffusion posterior sampling and then the inverse transform is applied to generate clean 3D point cloud. The method is very similar to the R2DM (Nakashima 2024) paper which is also used as the pretrained diffusion model in this work. The results for the R2DM method is not given in the paper though, which is strange.

Critical strengths

None

Critical weaknesses
  1. The paper makes incremental contribution to the R2DM (Nakashima & Kurazume 2024) paper. The results from baseline R2DM is missing from the paper despite the fact that a pretrained R2DM model is used in this paper. I will like to see how the EMD addition improves the performance over R2DM.

  2. The comparison with DPS and DPS+TV is not sufficient. Also, it is not clear if the authors train the DPS model on KITTI dataset or not ? Since R2DM is pretrained on KITTI dataset the comparison with DPS can only be valid if DPS is also trained on KITTI.

  3. The ablation study in table 3 shows that there are other surrogate designs that perform better than EMD (atleast for depth and reflectance MAE). Also, its not clear why the point cloud and range error metrics (FPD and FRD) are high for SSIM when depth MAE is low ?

Questions for authors

Questions:

  1. Can you clarify if the naive DPS model (which is used for comparison) was trained on KITTI dataset ?
  2. In table 3 the ablation study show that there are other surrogate design that work better than EMD in some cases. Can you provide more discussion on this?
  3. The forward operator (A) relies heavily on exact, test-time oracle metadata, such as identical lidar ring drop IDs or precise synthetic noise maps. How will this framework perform in real-world scenarios where the exact degradation matrix and parameters are completely unknown?

Suggestion:

  1. Test on real-world adverse weather data (e.g. Canadian Adverse Driving Conditions dataset) and show how the method translates to real-world degradations.
5 comments
AuthorsOpenReview ↗

Official Comment for Reviewer 7azv (1/2)

We thank the reviewer for the detailed reading of our setup. The three weaknesses point at places where our writing left the role of R2DM, the identity of the DPS baseline and the reading of Table 3 unclear, and we respond to each below.

W1. R2DM is our prior, not a baseline.

We believe there is a misunderstanding here, and we would like to clarify it. R2DM is not a method for LiDAR inverse problems, and our method is not a retrained or modified R2DM. R2DM is an unconditional generative model of clean LiDAR scans, and we use it as the prior with which we solve an inverse problem, that of recovering the clean scene behind a corrupted measurement yy. R2DM itself never takes yy, so it is not a baseline for that problem, and this is why no R2DM row appears in the paper.

Our paper supplies what R2DM lacks, namely the connection between a scene and a measurement of it, as differentiable forward operators for eight 3D corruptions and the correlative guidance built on the likelihood they define. The weights and the architecture of R2DM are untouched and we train nothing.

The comparison the reviewer asks for is already in Table 1 of the paper, where every diffusion row uses R2DM as its prior. The DPS row is R2DM with the standard ℓ2\ell_2 guidance, and the Ours row is R2DM with that guidance and our correlative term, under one checkpoint, one forward operator and 128 sampling steps. The two rows differ in the EMD term alone, so reading them against each other is reading exactly what the EMD addition contributes, and Table 1 gives it for all eight corruptions.

W2, Q1. The DPS baseline.

We would like to clarify how the DPS baseline is set up. Nothing is trained here. DPS is a sampling procedure rather than a model, which takes a pretrained diffusion prior and adds a measurement-likelihood gradient at each reverse step, so there is no DPS model that could have been trained on any dataset. Every row of our comparison runs on one and the same diffusion model, the R2DM pretrained on SemanticKITTI, and on KITTI-360 on the R2DM pretrained on KITTI-360, as stated in Section 5.4. DPS and our method thus share the prior and its training data, and differ in the guidance alone.

W3, Q2. The surrogate ablation.

We are glad to discuss Table 3 of the paper in more detail, and we will add this discussion to the paper. Two alternatives each lead on a single metric, SSIM on depth MAE and histogram intersection on BEV JSD, and in both cases by a fraction of a percent, while the same two rows fall far behind on FPD. EMD leads on the remaining metrics and ties for the best reflectance MAE.

Why SSIM can hold a low depth MAE and a high FPD at the same time follows from what the two measure. Depth MAE averages over every pixel, and under Beam Missing about two thirds of the pixels lie on the observed support, where the measurement already fixes the value, so the average is dominated by the region where little is at stake. FPD and FRD compare feature distributions of the reconstructed point cloud and range image, so they respond to whether the recovered surfaces sit where surfaces can sit. SSIM matches local means, variances and covariance inside a window, and a patch that spreads a depth discontinuity over several pixels preserves those statistics while placing points between two surfaces that belong to neither. Such points lower the average absolute error and are what a point-distribution metric charges for.

Sorted-L1 shows the same effect from the other side, and it also explains our choice within the distribution-matching family. Comparing sorted depths in a patch is the same assignment-invariant comparison as EMD, but its cost grows without bound in the displacement, so a single gross outlier in a patch dominates the term. The binned soft histogram in the log-depth domain bounds that cost, and that is where the two rows separate on the distributional metrics.

Since the deliverable of this task is a point cloud that a downstream model consumes, we select the surrogate on the distributional metrics and treat per-pixel MAE as secondary, because a surrogate that pulls every pixel toward a local consensus lowers the average while flattening the geometry. We will add this to Section 5.4 and report depth MAE split over the observed and unobserved supports alongside the average. We thank the reviewer for raising this point.

AuthorsOpenReview ↗

Official Comment for Reviewer 7azv (2/2)

Q3. Unknown operators.

We fully agree that solving LiDAR inverse problems when the exact degradation is unknown is an important direction, and we would like to clarify where our work stands on this. In the image domain the blind setting, where the operator is unknown and is recovered along with the signal, is already being taken on directly, as in BlindDPS. Such a method still evaluates the operator it estimates inside a likelihood, so the likelihood has to be right first, and for range-view LiDAR that is what was missing. We hope our work can be read as the groundwork that a blind treatment of LiDAR would build on.

We would also note what the operator is given in the meantime. Every entry it reads states which directions carry a return, how far the corruption displaced them, or by how much it attenuated or rescaled them, and none of them states what the clean scan held there. Knowing a realization does not undo it either, since the projection is not invertible and what the corruption removed stays in the null space of A\mathcal{A}.

Section 5.5 already reports one case with an approximate operator, a fog-like corruption that suppresses points within 6 m, and we have extended it into a quantitative study. Here A^\hat{\mathcal{A}} is estimated from the measurement alone, from the valid support of yy and a local-depth inlier set, without the corruption type or its realization. Both methods receive the same A^\hat{\mathcal{A}} and the guidance weights were not re-tuned.

Table x1. Performance when the forward operator must be estimated from the measurement alone. N=1,018N = 1{,}018. Lower is better for every metric.

CorruptionMethodFPD ↓FRD ↓JSD ×10² ↓MMD ×10⁴ ↓MAE (m) ↓
SnowDPS162.2531.52.3940.3760.519
SnowOurs120.3514.52.2400.3550.510
CrosstalkDPS12.33324.31.5040.2000.297
CrosstalkOurs8.20312.41.4380.1850.290
FogDPS23.03308.94.1350.2350.675
FogOurs21.10300.84.1050.2290.666

Every one of the five metrics favours our method on all three corruptions, so the correlative guidance remains useful when the exact operator is not available.

In the revision we will put the study above into the main paper in place of Section 5.5, discuss BlindDPS in the related work and add the blind setting as future work. We thank the reviewer for this question.

S1. Real adverse weather.

We agree that evaluation on real adverse weather matters, and we have run our method on it. Following the suggestion we applied the same observation-derived A^\hat{\mathcal{A}} to the Canadian Adverse Driving Conditions dataset, recorded in real snowfall, with no degradation parameters of any kind.

Neither the operator nor the prior is matched to this data. The operator is read off the measurement, and the prior is the same R2DM checkpoint used throughout the paper, trained on SemanticKITTI, so it was fit to a sensor with twice the beams, in a different city, and it has never seen a snowy scan. Nothing was retrained or fine-tuned.

Even so, the airborne returns that fill the near field of the measurement are removed, and the ground rings and the structures beside the road come back continuous. Real snowfall has no clean counterpart to score against, so we report this qualitatively. The reconstructions are now in Appendix G.3 of the revised supplementary material. We thank the reviewer for the suggestion.

AuthorsOpenReview ↗

Official Comment for Reviewer 7azv

Dear Reviewer 7azv,

We would like to thank you once again for your review and for your suggestions, which led us to run a study with the forward operator estimated from the measurement alone and to evaluate our method on real snowfall.

We would very much like to resolve the concerns you raised. Some of them, in particular the role of R2DM in our framework and the way the DPS baseline is set up, we believe our writing left unclear, and we have tried to clarify both in our response. We would be grateful for the chance to address anything that remains, and we would gladly run additional experiments while the discussion period is open.

Reviewer_12KaOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper adapts diffusion posterior sampling to LiDAR reconstruction using a pretrained range-view diffusion prior. It constructs differentiable corruption operators and augments pixel-wise L2 guidance with patch-wise Earth Mover’s Distance between local depth distributions. Experiments evaluate eight simulated corruptions on SemanticKITTI, with additional tests on KITTI-360, supervised denoisers, and alternative guidance terms.

Critical strengths
  1. The framework handles several corruption types without training a separate restoration model for each corruption.
  2. Table 1 shows substantial FPD reductions for several corruptions, together with improvements in other metrics. Evaluation on 4,071 SemanticKITTI frames provides meaningful coverage.
  3. The appendices document corruption metadata, histogram construction, guidance parameters, and runtime. EMD adds less than 1% runtime relative to the underlying DPS sampler.
Critical weaknesses
  1. A diagonal Gaussian likelihood assumes that measurement errors are conditionally independent given the clean signal; it does not require the clean signal itself to have independent pixels. Spatial dependencies can already be represented by the diffusion prior. Therefore, adjacent LiDAR points belonging to the same surface does not, by itself, establish that the likelihood requires a non-diagonal noise covariance. The paper needs evidence of correlated measurement residuals to support this argument. This distinction is consistent with the measurement formulation in DPS.
  2. Overlapping patches and L2 guidance can still provide useful constraints, but the retrieval experiment only establishes descriptor usefulness, not equivalence to a covariance correction. The method is better supported as an empirical local-distribution consistency penalty.
  3. The method records exact noise offsets, attenuation/scaling factors, and masks; Motion Blur additionally uses clean-to-corrupted correspondences and a clean XYZ grid. Knowing a stochastic corruption model is different from knowing its realized perturbations. Such information is generally unavailable from a real corrupted scan and is not described as being supplied to the supervised baselines. Consequently, the comparisons do not isolate restoration ability under matched information.
Questions for authors

see weaknesses

4 comments
AuthorsOpenReview ↗

Official Comment for Reviewer 12Ka (1/2)

We thank the reviewer for the careful and constructive review, and for the encouraging assessment of our work. We are glad that the reviewer found value in handling several corruptions without a separate restoration model for each, in the coverage of the evaluation, and in the documentation in the appendices. We are equally grateful for the precision of the three weaknesses raised. Each one names a specific gap between what we argue and what we show, which made clear what to measure and what to rewrite, and they have shaped our revision accordingly. We respond to each below and hope our responses resolve the concerns.

W1. Correlated measurement residuals.

We fully agree with the distinction. A diagonal Gaussian likelihood constrains the measurement error conditioned on the clean signal and places no requirement on the clean signal itself, the spatial dependencies of the scene are carried by the diffusion prior, and adjacent points lying on the same surface therefore does not by itself establish that the likelihood needs a non-diagonal covariance. Our paper points to the geometry of LiDAR range views but does not carry it through to the measurement noise. We would like to supply both that step and the evidence the reviewer asks for.

The noise nn in Eq. 12 is the part of the measurement that A(x0)\mathcal{A}(x_0) does not account for. A range view is a 2.5D projection, so it keeps only the nearest return along each ray, while the corruptions act on the full 3D scene. Consider a continuous 3D surface observed across neighboring rays. Corruptions can remove, displace, or replace returns from this surface before projection. For example in Snow, where snow particles can replace the original surface returns. The projection retains only the resulting returns, potentially losing geometric relationships between the original surface and the corrupted measurements. This information loss is inherent to the range-view representation, even when the corruption parameters are precisely known. Since these geometric relationships span neighboring rays, the discrepancies that remain in nn can also exhibit spatial dependencies rather than arise independently at each pixel. These residuals therefore carry spatial relationships that are not necessarily eliminated by conditioning on the clean range view, and the prior cannot represent them either.

We then measured the conditional quantity directly. Holding one scene fixed and redrawing only the corruption, we computed the autocorrelation of n=y−A(x0)n = y - \mathcal{A}(x_0) across draws, removing the mean over draws at each pixel. The mean over draws is what the scene fixes, so what the estimator returns is the conditional covariance Cov(n∣x0)\mathrm{Cov}(n \mid x_0), which is the quantity the diagonal assumption concerns. For reference we also report the spatial autocorrelation within a single draw, where the scene component is still present.

Table x1. Autocorrelation of the measurement residual under Beam Missing at lag kk, over 32 frames with 32 corruption draws per frame. The conditional rows hold the scene fixed and remove the mean over draws at each pixel. The single-draw rows keep the scene component.

Axisk=1k=1k=2k=2k=4k=4k=8k=8
Conditional on x0x_0azimuth+0.777+0.668+0.543+0.406
Conditional on x0x_0elevation+0.001−0.020−0.014−0.005
Single drawazimuth+0.756+0.645+0.518+0.376
Single drawelevation+0.080+0.043+0.099+0.059

The residual stays correlated after the scene component is removed, and along azimuth it is no weaker than before the removal, +0.777 against +0.756 at lag 1. What the prior could represent has been taken out, and the correlation remains.

The anisotropy identifies its source. A ring is removed by a single draw that is independent of the scene and that decides every column of the affected row at once, which couples a row completely and leaves rows uncoupled, and the measurement shows exactly that, +0.777 along azimuth against +0.001 along elevation.

We will add this measurement to the appendix and the argument above to the introduction and to Section 4.3. We thank the reviewer for pressing on this point.

AuthorsOpenReview ↗

Official Comment for Reviewer 12Ka (2/2)

W2. What the retrieval experiment shows.

We fully agree that the retrieval experiment shows the local depth histogram to be informative about local geometry without establishing equivalence to a covariance correction. We do not claim that equivalence, and estimating Koff\mathcal{K}_{\mathrm{off}} would require a full posterior covariance over the range view at every sampling step.

What we claim is narrower and concerns the form of the penalty. An off-diagonal term changes the guidance in two ways, and we chose a surrogate that reproduces them.

One surface produces the residuals over a whole neighborhood, and a diagonal term reads them as that many separate measurements. The guidance then follows a single piece of evidence once per pixel. A penalty defined over the patch follows it once.

The measurement does not fix which pixel of the patch receives which depth. The surface that takes the place of a removed return is not recorded in the range view, so what the measurement fixes is the local distribution of depths. A penalty that compares distributions asks for what the measurement carries. A per-pixel penalty asks the estimate to match an assignment it does not.

Patch-wise EMD has both properties. We offer it not as the only surrogate with these properties, but as one that has them, and Appendix C adds that it tells local geometry apart, with a Precision@20 of 0.463 against 0.235 for absolute depth and 0.013 for random retrieval.

In Section 4.4 we will present the term as a local-distribution consistency penalty whose design follows these two properties, and state what the retrieval experiment does and does not establish. We thank the reviewer for the characterisation.

W3. Matched information.

We fully agree that knowing a stochastic corruption model is not the same as knowing its realized perturbations.

We are sorry that Appendix F left this ambiguous, since it lists what the simulator preserves rather than what the operator reads. The operator receives a description of the corruption and never the clean scan. Every entry it reads states which directions carry a return, how far the corruption displaced them, or by how much it attenuated or rescaled them, and none of them states what the clean scan held there. For Motion Blur in particular the clean XYZ grid is recorded but is not an input, and deleting it from the record leaves A\mathcal{A} unchanged. Knowing a realization does not undo it either, since the projection is not invertible and what the corruption removed stays in the null space of A\mathcal{A}.

The reviewer is right that the realized parameters of the weather corruptions are not available from a real scan, and we will scope the claims accordingly. We also report a study in which the operator is estimated from the measurement alone, in our reply to Reviewer ov9o, where our method keeps its advantage on all five metrics.

Every diffusion row uses the same checkpoint, the same records, the same operator and the same number of sampling steps, so the only difference between the DPS row and ours is the guidance term. The supervised denoisers are given knowledge of the corruption as well, through paired training data for it, while our method trains on nothing.

We will state in Appendix F which entries the operator reads, and what each method is given in Section 5.3. We thank the reviewer for raising this point.

AuthorsOpenReview ↗

Official Comment for Reviewer 12Ka

Dear Reviewer 12Ka,

We are deeply grateful for your review and for your encouraging assessment of our work, and we thank you sincerely for the time you gave it. It means a great deal to us that you found value in what we set out to do. Your comments pointed to the places where our paper would benefit from fuller explanation, and we hope the additional detail and measurements in our response answer them.

We have posted our responses and hope that they resolve your concerns. If anything remains unclear or unconvincing, we would be very glad to discuss it further while the discussion period is open, and we would gladly run additional experiments if that would help.

AuthorsOpenReview ↗

Official Comment for Reviewer 12Ka

Dear Reviewer 12Ka,

As the discussion period comes to a close, we would like to thank you once again for your review and for your generous assessment of our work.

We are grateful too for the points you raised. They were the ones that sharpened the paper most, and it is better for them. Should anything further be of help to you, we would be glad to provide it.

Thank you again for your time and for the encouragement your review gave us.

Reviewer_ov9oOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses a pretrained LiDAR diffusion model to reconstruct point clouds affected by different corruptions. It extends DPS with a patch-wise EMD loss that matches local depth distributions between the observation and the predicted measurement. The method is evaluated on eight corruption types, with improvements over DPS on most reported metrics.

Critical strengths
  • The method is simple and can reuse a pretrained model without corruption-specific retraining. The added guidance introduces little computational overhead.
  • The experiments cover several corruption types and show substantial gains in FPD in most cases.
  • The ablations compare alternative losses, patch sizes, and depth transformations, providing useful support for the design choices.
Critical weaknesses
  • The evaluation relies on detailed corruption information. Appendix F describes recording exact corruption realizations and clean-to-corrupted correspondences; the motion blur setup also records a clean coordinate grid. Such information is usually unavailable in practice. This does not necessarily make the comparison with DPS unfair, but it limits what the results establish about real-world performance.
  • The baseline comparison is incomplete. R2DM already supports RePaint-based completion, which is a relevant baseline for beam missing and subsampling. Also, excluding linear solvers across all tasks is not justified when some operators are simply fixed masks.
  • The explanation for the guidance is not fully supported. Spatial correlations in the clean scene do not by themselves imply correlated measurement noise. Local histogram matching may be useful, but the paper does not establish its claimed connection to the off-diagonal precision matrix.
Questions for authors
  • Which recorded corruption details are used during inference, and how is the clean coordinate grid used for motion blur? Please report quantitative results when the operator must be estimated or contains errors.
  • Please compare your method with R2DM’s RePaint-based completion under the same pretrained model on beam-missing and subsampling tasks.
  • What evidence connects the EMD guidance to the proposed noise-correlation model? Could its benefits instead come from a useful local distribution-matching loss?
8 comments
AuthorsOpenReview ↗

Official Comment for Reviewer ov9o (1/3)

We thank the reviewer for the careful review. We appreciate the reviewer finding value in reusing a pretrained model without corruption-specific retraining and in the breadth of our evaluation, and the three weaknesses raised are helpful feedback. We respond to each below and hope our responses resolve the concerns.

W1, Q1. Operator inputs.

We are sorry that Appendix F left this ambiguous. The operator receives a description of the corruption and never the clean scan. Every entry it reads states which directions carry a return, how far the corruption displaced them, or by how much it attenuated or rescaled them, and none of them states what the clean scan held there. For Motion Blur in particular the clean coordinate grid is recorded but is not an input, and deleting it from the record leaves A\mathcal{A} unchanged.

Knowing a realization does not undo it either. The operator applies the recorded degradation to the estimate and re-projects, and the projection is not invertible, so the measurement cannot be reverted by subtraction and what the corruption removed stays in the null space of A\mathcal{A}. Every method we compare receives the same operator, the same records, the same pretrained prior and the same number of sampling steps, and only the guidance differs.

We would also like to clarify where our work stands. In the image domain the blind setting, where the operator is unknown and is recovered along with the signal, is already being taken on directly, as in BlindDPS. Such a method still evaluates the operator it estimates inside a likelihood, so the likelihood has to be right first, and for range-view LiDAR that is what was missing. We hope our work can be read as the groundwork that a blind treatment of LiDAR would build on.

The reviewer is right that some of what the operator reads is not available in practice. The realized per-point attenuation and range-scaling parameters of the weather corruptions cannot be obtained from a real scan, so we replaced the oracle operator with one estimated from the measurement alone. A^\hat{\mathcal{A}} is the intersection of the valid support of yy with a local-smoothness inlier set, obtained by iterated trimming against the local median depth in an 11×1111 \times 11 neighbourhood. It assumes only that an unknown corruption removes some returns and turns others into gross outliers, it does not know which corruption it faces, and it is the same for every corruption type. Both methods receive the same A^\hat{\mathcal{A}} and the guidance weights were not re-tuned.

Table x1. Performance when the forward operator must be estimated from the measurement alone. N=1,018N = 1{,}018. Lower is better for every metric.

CorruptionOperatorMethodFPD ↓FRD ↓JSD ×10² ↓MMD ×10⁴ ↓MAE (m) ↓
SnowOracle A\mathcal{A}DPS31.02300.91.7980.2350.385
SnowOracle A\mathcal{A}Ours18.23285.61.6560.2170.377
SnowA^\hat{\mathcal{A}}DPS162.2531.52.3940.3760.519
SnowA^\hat{\mathcal{A}}Ours120.3514.52.2400.3550.510
CrosstalkOracle A\mathcal{A}DPS17.81184.51.1290.03930.207
CrosstalkOracle A\mathcal{A}Ours10.99170.01.0010.03220.199
CrosstalkA^\hat{\mathcal{A}}DPS12.33324.31.5040.2000.297
CrosstalkA^\hat{\mathcal{A}}Ours8.20312.41.4380.1850.290
FogOracle A\mathcal{A}DPS10.20208.61.5950.03340.291
FogOracle A\mathcal{A}Ours10.30202.01.5660.03460.284
FogA^\hat{\mathcal{A}}DPS23.03308.94.1350.2350.675
FogA^\hat{\mathcal{A}}Ours21.10300.84.1050.2290.666

Replacing the oracle operator costs both methods a great deal, as it must. The ordering is unchanged and every metric still favours our method on all three corruptions.

AuthorsOpenReview ↗

Official Comment for Reviewer ov9o (2/3)

We also ran the case where the operator is wrong rather than estimated. For a mask-type operator only one direction of error has any effect, since wrongly declaring an unobserved pixel observed leaves the data term with no return there, whereas wrongly declaring an observed pixel unobserved makes the solver discard that measurement. We perturb in that direction alone on Beam Missing, leaving yy untouched, by randomly discarding 1% and 5% of the observed pixels and by discarding two entire range-view rows. Both methods receive the identical wrong operator.

Table x2. Operator error on Beam Missing. N=1,018N = 1{,}018. yy is identical across all rows and only the operator is corrupted.

Operator errorMethodFPD ↓FRD ↓JSD ×10² ↓MMD ×10⁴ ↓MAE (m) ↓
exactDPS20.12525.51.7110.20980.5630
exactOurs9.98509.21.5970.19310.5623
flip 1%DPS19.53528.21.7170.20980.5655
flip 1%Ours9.91511.31.6090.20150.5628
flip 5%DPS20.31535.11.7370.21700.5766
flip 5%Ours10.09517.21.6920.22530.5679
2 rowsDPS20.80546.01.7450.22050.5847
2 rowsOurs10.89527.51.6530.21220.5826

Neither method is sensitive to error of this magnitude. Structured error costs our method relatively more, since discarding two whole rows raises our FPD by 9.1% against 3.4% for DPS and narrows the margin from 2.02× to 1.91×, which follows from the correlative term operating on local neighbourhoods. At the 5% level one metric favours DPS.

In the revision we will replace the qualitative proof-of-concept of Section 5.5 with the two studies above, give the list of entries the operator reads separately from the list the simulator preserves in Appendix F, and add the blind setting as future work. We thank the reviewer for raising this point.

W2, Q2. RePaint and linear solvers.

We thank the reviewer for pointing to RePaint. It is a relevant baseline wherever the forward operator is a fixed mask, and we have run it under the same pretrained model on Beam Missing and on Cross Sensor, which is the subsampling setting the reviewer refers to.

The comparison has to control for one thing. RePaint replaces the observed pixels with the measurement at every reverse step and so reproduces the measurement there almost exactly, and two thirds of the pixels lie on the observed support, so the aggregate metrics are dominated by a region that RePaint copies rather than reconstructs. We therefore also report our method with a single data-consistency projection, which applies that same substitution once to the final x^0\hat{x}_0.

Table x3. Comparison with RePaint-based completion under the same pretrained model. N=1,018N = 1{,}018.

CorruptionMethodFPD ↓FRD ↓JSD ×10² ↓MMD ×10⁴ ↓MAE (m) ↓rMAE ×10 ↓
Beam MissingRePaint7.164625.91.6140.28730.53430.416
Beam MissingOurs7.697524.91.7290.28850.56270.379
Beam MissingOurs + proj.3.489412.61.1140.11440.43490.363
Cross SensorRePaint19.979900.02.4380.62230.73130.754
Cross SensorOurs13.735637.12.2970.47450.66050.683
Cross SensorOurs + proj.8.715570.31.7830.29680.61130.676

On Cross Sensor our method is ahead of RePaint on every metric with no projection at all. On Beam Missing it is ahead of our method on four of the six metrics, and with the treatment of the observed support matched our method leads on every one of them, by 51.3% on FPD, 34.1% on FRD, 31.0% on BEV JSD, 60.2% on BEV MMD, 18.6% on depth MAE and 12.7% on reflectance MAE. What the aggregate rewarded on Beam Missing was the region where the measurement is already available.

The two approaches also differ in scope. RePaint applies where the forward operator is a pure mask, which covers three of the eight corruptions we study, whereas our framework addresses all eight with the same pretrained prior and no retraining by changing only the guidance. Under Fog, Snow, Crosstalk and Motion Blur the measurement is not a subset of the clean scan and there is no known region to hold fixed, so the substitution that makes RePaint strong here is not available.

On linear solvers, for a fixed mask the operator is linear with a non-trivial null space, and the minimum-norm linear solution is the measurement on the observed support and zero elsewhere, so it performs no reconstruction precisely where reconstruction is needed. We will state this and report the corresponding row.

We will add RePaint and the projected rows to the baselines in the revision. We thank the reviewer for the suggestion.

AuthorsOpenReview ↗

Official Comment for Reviewer ov9o (3/3)

W3, Q3. The likelihood and the EMD surrogate.

We fully agree that spatial correlations in the clean scene do not by themselves imply correlated measurement noise, and that a diagonal likelihood constrains the measurement noise conditioned on the clean signal rather than the clean signal itself. Our paper points to the geometry of LiDAR range views but does not carry it through to the measurement noise, and we would like to expand on this point here.

The noise of Eq. 12 is the part of the measurement that A(x0)\mathcal{A}(x_0) does not account for. A range view is a 2.5D projection, so it keeps only the nearest surface along each direction, while the corruptions act on the full 3D scene. Consider a continuous 3D surface observed across neighboring rays. Corruptions can remove, displace, or replace returns from this surface before projection. For example in Snow, where snow particles can replace the original surface returns. The projection retains only the resulting returns, potentially losing geometric relationships between the original surface and the corrupted measurements. This information loss is inherent to the range-view representation, even when the corruption parameters are precisely known. Since these geometric relationships span neighboring rays, the discrepancies that remain in nn can also exhibit spatial dependencies rather than arise independently at each pixel. Because these geometric relationships are not fully preserved in the clean signal, fixing the clean signal does not necessarily remove this dependence. A diagonal covariance denies exactly this, so we model the likelihood with a covariance that couples neighboring pixels.

On the connection to the off-diagonal precision, we do not claim that the patch-wise EMD reconstructs Koff\mathcal{K}_{\mathrm{off}}, which would require a full posterior covariance over the range view at every sampling step. An off-diagonal term changes the guidance in two ways, and a surrogate is principled to the extent that it reproduces them.

Corruption of one surface produces the residuals over a whole neighborhood, and a diagonal term reads them as that many separate measurements. The guidance then follows a single piece of evidence once per pixel. A penalty defined over the patch follows it once.

The measurement retains the resulting depth at each pixel but not the full geometric relationships between the original surface and the corrupted returns. When neighboring returns from a common surface are affected, their shared geometric structure is better represented collectively through a local depth distribution than by isolated pixel values. A penalty that compares distributions asks for what the measurement carries. A per-pixel penalty instead treats these related discrepancies independently.

Patch-wise EMD has both properties. We offer it not as the only surrogate with these properties, but as one that has them.

The reviewer asks whether the benefit could instead come from a useful local distribution-matching penalty. On our account the two are not alternatives, since the two properties above are the reason a penalty over local distributions is the form the off-diagonal term calls for. Such a penalty is of use only if it tells local geometry apart, and this is what Appendix C measures. Retrieval by the local depth histogram returns pixels from the same object instance and with the same surface topology, while retrieval by absolute depth follows iso-depth contours and mixes in unrelated background at the same range. The Precision@20 is 0.463 against 0.235, with 0.013 for random retrieval.

We will state the argument above in the introduction and in Section 4.3, and we will relabel the term in Section 4.4 together with what it does and does not claim. We thank the reviewer for pressing on this point.

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

The manuscript presents a diffusion framework designed to solve LiDAR inverse problems across various measurement corruptions. The authors adapt Diffusion Posterior Sampling (DPS) to range view representations of LiDAR data by leveraging a pretrained diffusion model. The primary contribution involves a correlative guidance term. This term uses patchwise Earth Mover Distance (EMD) between local depth histograms to supplement standard pixelwise L2 guidance. The authors evaluate the technique on SemanticKITTI across several distinct corruption types and include an experiment on KITTI 360. They provide comparisons with supervised LiDAR denoisers and present ablation studies.

Critical strengths
  1. The work addresses a relevant problem driven by practical motivation. Framing distinct LiDAR corruptions within one unified inverse problem framework enables the use of a pretrained generative model without requiring retraining for specific corruptions. The introduced patchwise EMD term integrates smoothly into DPS. The experiments indicate that this term supplies valuable additional measurement guidance.

  2. The experimental setup is broad, covering eight corruption types and multiple complementary metrics. That is, table 1 shows the technique improving FPD over the better performing model between DPS and DPS+TV across seven out of the eight corruptions. The researchers also test alternative surrogate losses, analyze heavy severity Beam Missing, and investigate patch sizes and bin transformations.

Critical weaknesses

1.I am concerned about the statistical justification provided for the correlative guidance. The core argument states that standard L2 guidance assumes spatial independence. To be concrete, the paper claims this assumption fails because adjacent LiDAR measurements link to correlated 3D geometry. However, the independence of measurement noise conditioned on the clean signal remains entirely distinct from the independence of the clean signal itself. For example, a diagonal measurement noise covariance does not automatically mean adjacent LiDAR depths act independently. Can the authors expand on this point?

  1. The experiments indicate the patchwise EMD acts as a helpful regularizer. Yet, viewing the gradient of this EMD as a principled approximation for the missing likelihood component demands further justification. The paper should also discuss the relationship between this approach and prior research focusing on structured noise diffusion models, like reference [1].

  2. The experimental evaluation raises questions regarding the available measurement information and how the reported improvements are attributed. Specifically, section 4.2 builds a differentiable forward operator utilizing fixed pixel correspondences alongside corruption maps., and Appendix F details the explicit use of exact corruption realizations and recorded metadata. For certain corruptions, this data includes realized noise fields and range scaling parameters. For Motion Blur, the construction goes further by recording clean pixel correspondences and a clean xyz grid.

  3. Assuming a known forward operator represents a standard practice in inverse problems. However, understanding the corruption mechanism does not equal knowing the precise stochastic corruption realization and the exact correspondence to the original clean scan. The manuscript, at least to my understanding, does not establishe how one might extract these quantities from actual corrupted measurements. I think it is important to clarify this point in the rebuttal.

  4. Recent LiDAR diffusion completion strategies, specifically references [2] and [3], are relevant but not discussed in the paper.

References:

[1] Removing Structured Noise using Diffusion Models

[2] Scaling Diffusion Models to Real World 3D LiDAR Scene Completion

[3] LiDPM: Rethinking Point Diffusion for Lidar Scene Completion

Questions for authors

Please see my requests in the review above.

5 comments
AuthorsOpenReview ↗

Official Comment for Reviewer gvUU (1/3)

We thank the reviewer for the careful review. We appreciate the reviewer finding value in our unified formulation and broad evaluation, and the five weaknesses raised are helpful feedback. We respond to each below and hope our responses resolve the concerns.

W1. Measurement noise independence.

We fully agree that the independence of measurement noise conditioned on the clean signal is distinct from the independence of the clean signal itself. Our paper points to the geometry of LiDAR range views but does not carry it through to the measurement noise, and we would like to expand on this point here.

The diagonal model is reasonable when the noise is generated separately at each sensor element, as in the image problems it comes from. Our paper states that this fails for a range view because adjacent pixels take their values from one surface. Adjacent pixels also take their measurement noise from one surface, and they do so even once the clean signal is fixed.

The noise nn in Eq. 12 represents the part of the measurement that A(x0)\mathcal{A}(x_0) does not account for. A range view is a 2.5D projection that retains only the nearest return along each ray, while corruptions act on the full 3D scene. Consider a continuous 3D surface observed across neighboring rays. Corruptions can remove, displace, or replace returns from this surface before projection. For example in Snow, where snow particles can replace the original surface returns. The projection retains only the resulting returns, potentially losing geometric relationships between the original surface and the corrupted measurements. This information loss is inherent to the range-view representation, even when the corruption parameters are precisely known. Since these geometric relationships span neighboring rays, the discrepancies that remain in nn can also exhibit spatial dependencies rather than arise independently at each pixel. These residuals therefore carry spatial relationships that are not necessarily eliminated by conditioning on the clean range view. A diagonal covariance neglects such relationships, so we model the likelihood with a covariance that couples neighboring pixels. Carrying the 3D structure that the projection leaves out is the role of our correlative guidance.

We will state the argument above explicitly in the introduction and in Section 4.3 of the revision. We thank the reviewer for pressing on this point.

W2. The EMD surrogate.

We agree that this needs further justification. An off-diagonal term changes the guidance in two ways, and a surrogate is principled to the extent that it reproduces them.

One surface produces the residuals over a whole neighborhood, and a diagonal term reads them as that many separate measurements. The guidance then follows a single piece of evidence once per pixel. A penalty defined over the patch follows it once.

The measurement does not fix which pixel of the patch receives which depth. The surface that takes the place of a removed return is not recorded in the range view, so what the measurement fixes is the local distribution of depths. A penalty that compares distributions asks for what the measurement carries. A per-pixel penalty asks the estimate to match an assignment it does not.

Patch-wise EMD has both properties. It must also tell local geometry apart, since a penalty that cannot would act only as a smoothness regularizer, and this is what Appendix C measures. Retrieval by the local depth histogram returns pixels from the same object instance and with the same surface topology, while retrieval by absolute depth follows iso-depth contours and mixes in unrelated background at the same range. The Precision@20 is 0.463 against 0.235, with 0.013 for random retrieval. We offer patch-wise EMD not as the only surrogate with these properties, but as one that has them.

Reference [1] shares our starting point, that an i.i.d. Gaussian likelihood is the wrong model once the noise has structure, and differs in where that structure is placed. It learns a second score model for the noise distribution and recovers signal and noise together through a joint conditional reverse diffusion, so the structure enters as a learned prior. This asks the noise to be a source with a distribution of its own, one that can be sampled and collected into a training set. The residual we face is the scene geometry that the range view does not store, which varies with the scene and with the corruption, so a noise prior would have to be learned for each one. Placing the structure in the covariance of the likelihood instead leaves the pretrained prior fixed and lets the corruptions differ only in the guidance.

We will add this discussion to the related work and to Section 4.4 of the revision. We thank the reviewer for pointing us to [1].

AuthorsOpenReview ↗

Official Comment for Reviewer gvUU (2/3)

W3. Operator inputs.

We are sorry that Appendix F left this ambiguous. The operator receives a description of the corruption and never the clean scan. For Motion Blur it reads the recorded 3D displacement of the corruption, that is the global translation plus the per-point jitter, together with the correspondence mask. For the weather corruptions the corruption map of Section 4.2 holds the realized attenuation and range-scaling parameters. The clean xyzxyz grid is recorded but is not an input, and deleting it from the record leaves A\mathcal{A} unchanged.

Knowing a realization does not undo it. The operator applies the recorded degradation to the estimate and re-projects, and the projection is not invertible, so the measurement cannot be reverted by subtraction and what the corruption removed stays in the null space of A\mathcal{A}.

Every method we compare receives the same operator, the same records, the same pretrained prior and the same number of sampling steps, and only the guidance differs. The improvements we report are therefore attributable to the guidance.

Section 4.2 holds the correspondence fixed during sampling, because the 3D corruption is discrete while the projection uses a non-differentiable zz-buffer. This is an assumption of the same character as treating the observed support as given in a mask-type operator.

Appendix F lists what the simulator preserves without separating it from what the operator reads. We will give the two lists separately in the revision, together with the shared protocol above.

W4. Realizations in practice.

We fully agree that knowing the corruption mechanism is not the same as knowing the realized corruption. We would like to clarify where our work stands on this. In the image domain the blind setting, where the operator is unknown and is recovered along with the signal, is already being taken on directly, as in BlindDPS. Such a method still evaluates the operator it estimates inside a likelihood, so the likelihood has to be right first, and for range-view LiDAR that is what was missing. We hope our work can be read as the groundwork that a blind treatment of LiDAR would build on.

Even so, we have asked how far the framework carries when the realization is not available. Section 5.5 reports one such case, a fog-like corruption that suppresses points within 6 m, where posterior sampling with an estimated surrogate operator recovers the near-range geometry while the rest of the scene is preserved. It is a qualitative proof of concept, and we have now made it quantitative.

One way to approximate the operator is to read it off the measurement. We take A^\hat{\mathcal{A}} to be the intersection of the valid support of yy with a local-smoothness inlier set, obtained by iterated trimming against the local median depth in an 11×1111 \times 11 neighbourhood. It assumes only that an unknown corruption removes some returns and turns others into gross outliers, it does not know which corruption it faces, and it is the same for every corruption type. We ran the three weather corruptions under this operator and under the oracle operator, with the guidance weights unchanged between the two.

Table x1. Performance when the forward operator must be estimated from the measurement alone. N=1,018N = 1{,}018. Lower is better for every metric.

CorruptionOperatorMethodFPD ↓FRD ↓JSD ×10² ↓MMD ×10⁴ ↓MAE (m) ↓
SnowOracle A\mathcal{A}DPS31.02300.91.7980.2350.385
SnowOracle A\mathcal{A}Ours18.23285.61.6560.2170.377
SnowA^\hat{\mathcal{A}}DPS162.2531.52.3940.3760.519
SnowA^\hat{\mathcal{A}}Ours120.3514.52.2400.3550.510
CrosstalkOracle A\mathcal{A}DPS17.81184.51.1290.03930.207
CrosstalkOracle A\mathcal{A}Ours10.99170.01.0010.03220.199
CrosstalkA^\hat{\mathcal{A}}DPS12.33324.31.5040.2000.297
CrosstalkA^\hat{\mathcal{A}}Ours8.20312.41.4380.1850.290
FogOracle A\mathcal{A}DPS10.20208.61.5950.03340.291
FogOracle A\mathcal{A}Ours10.30202.01.5660.03460.284
FogA^\hat{\mathcal{A}}DPS23.03308.94.1350.2350.675
FogA^\hat{\mathcal{A}}Ours21.10300.84.1050.2290.666

Results under the estimated operator are worse than under the oracle operator, as expected. Every one of the five metrics still favours our method on all three corruptions.

In the revision we will put this study into the main paper in place of Section 5.5, discuss BlindDPS in the related work, and add the blind setting for range-view LiDAR as future work. We thank the reviewer for raising this point.

AuthorsOpenReview ↗

Official Comment for Reviewer gvUU (3/3)

W5. Related point-space methods.

We thank the reviewer for these pointers and agree that both are relevant.

Reference [2] runs diffusion directly in point space and completes a scene from a sparse scan with a point-wise local formulation. Reference [3] shows that the local approximation is not needed at scene scale and that a vanilla DDPM started from a suitably chosen intermediate step suffices. Both train a diffusion model for completion and condition it on the input scan.

Scene completion is one instance of the problems we address, the mask-type one, so the two lines meet there and differ in what is learned. References [2] and [3] contribute on the prior and the sampler, and their point representation is free of the 2.5D restriction discussed in W1. We keep an unconditional pretrained range-view prior fixed, train nothing, and contribute on the likelihood, which is what lets a single formulation cover eight corruptions rather than one task.

We will add this discussion to Section 2 of the revision.

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your time and effort in reviewing this paper. The authors have posted responses to your reviews. If you have not yet done so, please read their responses and leave a follow-up comment visible to the authors, indicating which concerns have been addressed and which remain unresolved.

Please also update your review and rating if your assessment has changed. Your feedback will help the authors address any remaining questions during the discussion period.

Thank you!

Dear Reviewers gvUU, ov9o, 12Ka, and 7azv,

We sincerely thank all reviewers for their valuable and insightful reviews. Your comments have helped us substantially improve the quality and completeness of our paper, and we have revised the manuscript accordingly and uploaded the revised version.

All changes made in response to the reviews are shown in red text, including figure and table captions. Most of the additions, including the further explanations, experiments, and analyses from our responses, have been made in the supplementary material. The exceptions are the additional related work and future work, which have been added to the main text.

We thank the reviewers again for their time and constructive feedback.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.