acceptodds
Under review as a conference paper at ICLR 2027

PRISM: Point Cloud Restoration for Impaired Single-Photon Data via Latent Point Flow Matching

Abstract

Single-photon imaging enables 3D sensing under low illumination and at long range, making point-cloud restoration essential for converting sparse, contaminated measurements into usable geometry. However, existing object-level datasets offer limited shape diversity and paired reference geometry for evaluating restoration under controlled acquisition conditions. Spatially non-uniform sampling, missing surface observations, and spurious detections further complicate recovery by making geometric evidence unevenly reliable. To address these challenges, we introduce SPORE, a physics-grounded simulated benchmark spanning 55 ShapeNet categories and three corruption regimes, constructed using ray-based acquisition and Poisson photon statistics. We also propose PRISM, a hierarchical latent flow-matching framework that selectively repairs unreliable local anchors and retains their reliability and correction information to guide subsequent global–local restoration. On SPORE, PRISM achieves the best results on three of four metrics under Mid corruption and all four under High corruption, reducing HD95 by 27.6% and improving [email protected] by 22.9% relative to the strongest competing result for each metric in the High regime. These findings support reliability-informed restoration as an effective approach to recovering precise geometry under severe simulated single-photon degradation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
impaired single-photon data, point cloud restoration, geometry-aware latent repair, global-local flow matching

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_5U5cOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper targets object-level restoration of single-photon LiDAR point clouds, which are sparse, incomplete and contaminated by background detections. It introduces SPORE, a simulated benchmark built from ShapeNet meshes with ray casting, Poisson photon statistics and three corruption levels, and PRISM, a restoration method in LION's hierarchical latent space. PRISM first estimates the reliability of each local latent anchor and repairs unreliable ones using reliable neighbors, then runs coupled global and local flow matching conditioned on the recorded repair states. On SPORE it ranks first on all four metrics under High corruption and is competitive under Low and Mid, with one qualitative example on real SPAD data.

Critical strengths
  1. The benchmark fills a real gap. Paired clean geometry with physically motivated single-photon corruption across 55 categories and three severity levels doesn't exist yet, and the simulator is documented in enough detail to be reproduced.
  2. The method is reasonably motivated and the ablation is thoughtfully designed.
Critical weaknesses
  1. The [email protected] results are hard to trust. The same model gets a higher [email protected] as the input gets worse (0.200, 0.214 and 0.342 for Low, Mid and High), and under Low it scores the same as the noisy input. This suggests the metric mostly reflects sampling alignment with the FPS-sampled reference rather than restoration quality, which matters because the main High-regime gains are on [email protected] and HD95.
  2. The ablation results are difficult to interpret. Removing a single conditioning component drops [email protected] from 0.342 to 0.05–0.07 while CD changes much less, so it's unclear what each module actually contributes.
  3. The real-world evaluation is limited to one qualitative example. Since the benchmark is simulated on single ShapeNet objects, the paper doesn't show that the method transfers to real single-photon data.
  4. The technical novelty is limited. The method mainly combines existing components, such as LION's latent space, flow matching and FiLM-style conditioning, and under Low corruption it is behind PGD and PD-LTS on CD and P2M.
Questions for authors
  1. Why does PRISM's [email protected] increase with corruption for the same model?
  2. The autoencoder is trained on clean FPS-sampled clouds. Do PRISM's outputs inherit that sampling pattern? If so, how does PRISM compare with methods that only displace input points under a sampling-insensitive metric?
  3. Can you provide quantitative real-data results, for example against a reference scan, including the strongest baselines?
  4. How sensitive is the method to the simulator settings?
18 comments
AuthorsOpenReview ↗

Response to Question 1

We appreciate the reviewer's careful observation regarding the [email protected] trend across corruption regimes. This behavior can be understood from three perspectives: PRISM's reliability-conditioned restoration mechanism, the precision–recall relationship, and the threshold sensitivity of the F-score.

Reliability-conditioned Restoration. PRISM is designed to address spatially heterogeneous and structured corruption in single-photon point clouds. GALR selectively repairs observations according to their estimated reliability, while the resulting repair states guide GLFM's global–local latent transport. Under Low corruption, the observed geometry is already relatively reliable, leaving limited room for further refinement. Moreover, some points may remain outside the [email protected] distance threshold even after their geometric errors are reduced. Consequently, improvements in geometric accuracy do not necessarily translate into higher F-scores at a fixed threshold. In contrast, stronger corruption provides greater scope for reliability-guided restoration of unreliable observations and missing structures.

Precision–Recall Analysis. Under Low corruption, precision increases from 0.2300 to 0.2500, while recall decreases from 0.1759 to 0.1663, leaving [email protected] nearly unchanged (0.1993 to 0.1997). In contrast, under High corruption, both precision (0.0700 to 0.4200) and recall (0.0465 to 0.2884) improve substantially, increasing [email protected] from 0.0559 to 0.3419. These results reveal a precision–recall trade-off under Low corruption, whereas severe corruption benefits from simultaneous improvements in correspondence accuracy and geometric coverage.

Threshold Sensitivity. The F-score evaluates point-cloud correspondence within a fixed distance threshold and is therefore sensitive to nearest-neighbor distance distributions. At a stricter threshold of 0.0075, F-scores follow Low > Mid > High, with values of 0.1650, 0.1450, and 0.1100, respectively. However, at 0.01, the ordering reverses, with scores of 0.1997, 0.2143, and 0.3419. Increasing the threshold from 0.0075 to 0.01 produces a much larger F-score gain under High corruption (0.2319) than under Low (0.0347) or Mid (0.0693). This indicates differences in nearest-neighbor distance distributions across corruption regimes, suggesting that some geometric improvements may not be fully reflected by a fixed-threshold F-score.

Importantly, this non-monotonic behavior is not unique to PRISM. As reported in Table 1, PD-Flow achieves [email protected] scores of 0.1306 and 0.2781 under Low and High corruption, respectively, while PCN achieves 0.1856 and 0.2293. Both methods obtain higher [email protected] under High corruption than under Low corruption, and neither exceeds the noisy input's [email protected] of 0.1993 under Low corruption. This further suggests that the observed trend is not specific to PRISM but is associated with differences in geometric error distributions and the threshold-dependent nature of the F-score.

Overall, the limited [email protected] improvement under Low corruption may result from the combined effects of limited corrective refinement, the precision–recall trade-off, and fixed-threshold evaluation. The higher [email protected] under High corruption does not imply uniformly lower geometric errors than under Low corruption, but instead reflects differences in the distributions of point-cloud correspondence distances.

AuthorsOpenReview ↗

Response to Question 2

We thank the reviewer for raising this important question. Farthest Point Sampling (FPS) is a well-established sampling strategy in point cloud processing, widely used to obtain spatially distributed representative points. For example, PointNet++ (Qi et al., 2017) employs FPS for hierarchical point sampling. Following this established practice, we used FPS to construct the original 2,048-point clean references for evaluation. Importantly, PRISM does not explicitly apply FPS to its restored outputs during inference. GALR selectively corrects local spatial anchors according to their geometric reliability, while GLFM performs repair-state-conditioned global–local latent transport. The decoder subsequently predicts coordinate residuals relative to the transported local anchors. Thus, the output coordinates arise from input-conditioned geometric restoration rather than explicit FPS resampling.

To assess the influence of reference sampling, we conducted an additional reference-resampling experiment. Keeping the denoised point clouds from all methods unchanged, we randomly sampled 2,048 reference points from the same underlying clean mesh surfaces to construct Random-2048 references and recomputed CD-L2, [email protected], and HD95. The results show that PRISM maintains the highest [email protected] and lowest HD95 across all three corruption regimes. Under High corruption, PRISM achieves a CD-L2 of 5.401, [email protected] of 0.3186, and HD95 of 0.0316, outperforming PGD and PD-LTS on all three metrics. Compared with the original FPS-2048 evaluation, its [email protected] decreases by 6.8%, comparable to the changes observed for PGD (6.2%) and PD-LTS (7.5%). These results indicate that PRISM's performance advantage persists when the reference sampling strategy changes.

Noise LevelMethodCD-L2 ↓P2M ↓[email protected] ↑HD95 ↓
LowPGD0.9740.9890.16380.0340
LowPD-LTS1.0291.1040.17060.0347
LowPRISM1.1981.2400.19040.0292
MidPGD1.2301.2000.13000.0363
MidPD-LTS1.2541.2010.12560.0384
MidPRISM1.2121.2420.20450.0248
HighPGD5.5332.6320.08440.0515
HighPD-LTS5.5283.5450.07120.0574
HighPRISM5.4012.1540.31860.0316

Our original bidirectional Point-to-Mesh (P2M) evaluation provides complementary evidence. P2M directly measures point-to-surface and surface-to-point geometric discrepancies against the ground-truth mesh, independently of the discrete reference sampling strategy. Therefore, the P2M values remain unchanged in the reference-resampling experiment. Under High corruption, PRISM achieves a P2M of 2.154, compared with 2.632 for PGD and 3.545 for PD-LTS, corresponding to error reductions of 18.2% and 39.2%, respectively. Together, the reference-resampling experiment and surface-based evaluation indicate that PRISM's improvements reflect geometric restoration beyond alignment with an FPS-specific reference distribution.

AuthorsOpenReview ↗

Response to Question 3 and Question 4

Q3: Thank you for this comment. We tested PRISM, PGD, and PD-LTS on eight real, experimentally acquired single-photon point-cloud samples. Since clean ground truth has not yet been obtained, we used two no-reference geometric descriptors, local surface variation (SV) and isolated-point ratio (IPR).

After removing exact duplicate points, SV was calculated using each point’s 20 nearest neighbors, excluding the query point itself. It is the ratio of the smallest neighborhood covariance eigenvalue to the sum of all eigenvalues, averaged over all points. Lower SV indicates lower local nonplanarity. IPR measures the percentage of points with fewer than three other neighbors within a specified radius. This radius was set to three times the median nearest-neighbor distance in the deduplicated noisy input and was shared by all methods within each sample. Lower IPR indicates fewer isolated points.

The following table reports the unweighted means across the eight samples. PRISM achieved the lowest IPR on 8/8 samples and the lowest SV on 6/8 samples. These results indicate that PRISM continues to outperform PGD and PD-LTS overall in terms of SV and IPR on the evaluated real single-photon point clouds, providing additional no-reference evidence of its denoising effectiveness in this setting. Detailed visualizations and analysis, together with the quantitative comparison, will be provided in Section 5.3, “REAL-WORLD RESULTS,” of the revised manuscript.

MetricMethodSample 01Sample 02Sample 03Sample 04Sample 05Sample 06Sample 07Sample 08Mean
SV ↓Noisy0.19690.11360.14860.19440.06750.07190.13580.06100.1237
SV ↓PGD0.15400.10270.11880.17280.05330.06970.13640.05510.1078
SV ↓PD-LTS0.18900.10650.14700.18570.05840.06370.12670.04320.1150
SV ↓PRISM0.18030.02540.09450.17540.02740.02850.12260.03310.0859
IPR (%) ↓Noisy12.999.585.2224.9230.7323.4723.0021.0818.87
IPR (%) ↓PGD1.323.540.390.3729.7719.2518.7820.2411.71
IPR (%) ↓PD-LTS6.786.681.6113.6426.8422.1714.9620.1514.10
IPR (%) ↓PRISM1.260.050.120.091.910.190.671.160.68

Q4: We thank the reviewer for raising this important question. To provide a systematic evaluation, we have added a dedicated section, “Robustness to Simulator and Acquisition Shifts,” to the revised Appendix. We conduct 20 Mid-anchored experiments, comprising one nominal reference, 12 single-factor perturbations, three combined corruption settings, and four acquisition-geometry variations. These experiments cover changes in photon emission, background noise, dropout, timing and angular jitter, spurious detections, viewpoint configurations, and field of view. All evaluations use the same trained PRISM model without additional fine-tuning.

The results demonstrate PRISM's consistent restoration capability across the tested parameter shifts. Across all 19 shifted conditions, PRISM improves all four evaluation metrics relative to the corresponding noisy inputs, reducing CD-L2 by 93.4–97.9% and P2M by 92.9–97.9%. For example, increasing the background rate from 0.15 to 0.60 raises the noisy CD-L2 from 32.813 to 66.762, while PRISM maintains a low CD-L2 of 1.421. Even under the combined low-photon and high-dropout condition, PRISM reduces CD-L2 from 180.318 to 8.318 and HD95 from 0.8230 to 0.0387. With only three acquisition views, it further reduces CD-L2 from 38.902 to 2.571 while improving [email protected] from 0.0833 to 0.2421.

These results provide quantitative evidence that PRISM maintains effective geometric restoration across diverse simulator configurations, including substantial changes in photon statistics, sensor noise, and acquisition geometry. The complete parameter settings, quantitative comparisons, and detailed analyses are now included in the updated Appendix.

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

This paper introduces SPORE, a simulated benchmark for restoring point clouds degraded by single-photon acquisition, and PRISM, a hierarchical latent flow-matching restoration framework. SPORE covers 55 ShapeNet categories and three corruption regimes, combining ray-based acquisition, Poisson photon statistics, missed returns, spurious detections, and measurement uncertainty. PRISM first performs Geometry-Aware Latent Repair (GALR), estimating token reliability and selectively correcting local spatial anchors. It retains reliability estimates and applied corrections as explicit repair states. Global–Local Flow Matching (GLFM) subsequently restores the hierarchical representation while using these states to condition local interactions and feature modulation. Experiments compare nine baselines retrained on SPORE. PRISM achieves the best results on three of four metrics under Mid corruption and all four under High corruption. Additional experiments examine component contributions, simulator consistency, and qualitative transfer to real measurements.

Critical strengths
  1. SPORE provides a substantial, carefully documented resource for studying structured corruption. The benchmark includes diverse object categories, explicit object-level splits, and shared evaluation pairs across methods. The appendix documents geometric checks, photon-count consistency, degradation progression before and after preprocessing, and comparisons against a mean-distance-matched Gaussian control. These analyses make SPORE useful as a reproducible, controlled restoration benchmark, even though its correspondence to real sensor statistics requires further validation.
  2. The ablations provide meaningful evidence that persistent repair information adds value beyond preprocessing alone. The repair-state-conditioning ablation retains the same repaired initialization while removing downstream GALR-derived conditioning. Under High corruption, [email protected] decreases from 0.3419 to 0.0718 and HD95 increases from 0.0297 to 0.0618. Additional ablations retain ordinary local message passing while removing reliability-dependent neighborhood processing or point-wise FiLM. Together with the baseline comparisons, these results support the usefulness of carrying repair information through subsequent restoration.
Critical weaknesses
  1. The physical representativeness of the benchmark is insufficiently established. The simulator aggregates ten viewpoints, generates background detections only on rays with valid object intersections, and uses uniform range replacement for background-only and spurious returns. These are consequential modeling choices. The geometric and Poisson checks establish implementation consistency, but do not establish agreement with measured acquisition statistics. Section 5.3 provides a qualitative real-capture example without quantitative reference geometry or baseline comparisons. The relationship between this capture and the ten-view full-body simulation is also unclear. Consequently, the evidence is considerably stronger for restoration under the specified synthetic corruption process than for realistic single-photon acquisition more broadly.
  2. The reliability estimator’s bounded calibration can prevent learned context from correcting an inappropriate geometric prior. Appendix D.2 sets the prior logit to (s_i-2)/0.25, while the two learned correction heads together can change it by at most 4. For an anchor with standardized spacing s_i=0, the predicted corruption probability therefore cannot exceed approximately 0.018, regardless of its latent context. This remains below the 0.25 threshold for any GALR correction. Thus, geometrically dense but erroneous clusters can be forced to receive high reliability and no pre-repair. GLFM may subsequently correct these errors, but the paper does not quantify this restriction’s practical effect. Figure 9 provides an illustrative visualization rather than dataset-level validation of reliability estimation. Since reliability controls both repair and subsequent information propagation, its behavior in structurally ambiguous regions deserves direct evaluation.
Questions for authors
  1. Which measured sensor statistics support the simulator’s background-range distribution, detection rules, and corruption parameters? How does the real acquisition in Figure 4 correspond to the simulated ten-view setting, and are quantitative reference measurements or baseline comparisons available?
  2. Is the bounded calibration behavior described above intentional? What fraction of geometrically inaccurate anchors receive low corruption probabilities because their local spacing is unexceptional, and how does this affect compact outlier clusters and weakly supported structures?
  3. How much does learned calibration improve reliability estimation and final restoration over the geometry-derived prior alone? A comparison against the training-time clean-aware targets would help establish what the learned context contributes.
5 comments
AuthorsOpenReview ↗

Response to Question 1

q1:We thank the reviewer for highlighting the need to clarify the simulator's empirical grounding and real-world validation. We address the questions regarding sensor statistics, acquisition settings, and quantitative evaluation in turn. q1:SPORE's degradation model is grounded in established statistical characterizations of single-photon LiDAR measurements and is further supported by comparisons with real measurements from SP-TransientBench.

Physical basis. Tachella et al. (2019) model photon detections using a Poisson observation process combining signal returns, instrumental temporal response, and background counts. Chen et al. (2020) similarly adopt a signal-plus-background Poisson model and characterize the challenges of photon-starved measurements and incomplete spatial observations. More recently, Kitichotkul et al. (2025) investigate how background flux, detector dead time, and acquisition protocols affect photon detection and depth estimation. These findings motivate SPORE's modeling of Poisson photon counts, visibility-dependent observations, stochastic missed returns, background-induced outliers, and ranging perturbations. In particular, the commonly adopted time-stationary background model motivates approximately uniform background arrival times within the acquisition window, which SPORE represents through uniform range replacement for background-dominated returns. The Low/Mid/High regimes provide controlled variations in degradation severity. Appendix B.4 further verifies the internal consistency of our Poisson sampling implementation.

Empirical validation. We additionally compare SPORE with real single-photon point clouds from SP-TransientBench. The median KNN spacing coefficient of variation is 0.721 for STB and 0.771 for High-corruption SPORE, with 98.8% of simulated samples falling within the observed STB range of 0.592–1.188. The median planarity is 0.425 for STB versus 0.457 for Mid-corruption SPORE; scattering is 0.174 versus 0.191 for High corruption; and linearity is 0.360 versus 0.336 for Low corruption.

Together, established photon-counting models and the observed overlap in local geometry and sampling statistics provide complementary physical and empirical support for SPORE's degradation design. They establish SPORE as a physically motivated benchmark for evaluating restoration robustness across controlled single-photon corruption conditions.

q2:Figure 4 illustrates the real single-photon acquisition system used to collect samples 01, 02, and 04 in our real-world dataset. The simulator's ten-view configuration models observations from multiple viewpoints, capturing viewpoint-dependent visibility, occlusion, and spatial sampling characteristics.

AuthorsOpenReview ↗

Response to Question 1

Thank you for this comment. We tested PRISM, PGD, and PD-LTS on eight real, experimentally acquired single-photon point-cloud samples. Since clean ground truth has not yet been obtained, we used two no-reference geometric descriptors, local surface variation (SV) and isolated-point ratio (IPR).

After removing exact duplicate points, SV was calculated using each point’s 20 nearest neighbors, excluding the query point itself. It is the ratio of the smallest neighborhood covariance eigenvalue to the sum of all eigenvalues, averaged over all points. Lower SV indicates lower local nonplanarity. IPR measures the percentage of points with fewer than three other neighbors within a specified radius. This radius was set to three times the median nearest-neighbor distance in the deduplicated noisy input and was shared by all methods within each sample. Lower IPR indicates fewer isolated points.

The following table reports the unweighted means across the eight samples. PRISM achieved the lowest IPR on 8/8 samples and the lowest SV on 6/8 samples. These results indicate that PRISM continues to outperform PGD and PD-LTS overall in terms of SV and IPR on the evaluated real single-photon point clouds, providing additional no-reference evidence of its denoising effectiveness in this setting. Detailed visualizations and analysis, together with the quantitative comparison, will be provided in Section 5.3, “REAL-WORLD RESULTS,” of the revised manuscript.

MetricMethodSample 01Sample 02Sample 03Sample 04Sample 05Sample 06Sample 07Sample 08Mean
SV ↓Noisy0.19690.11360.14860.19440.06750.07190.13580.06100.1237
SV ↓PGD0.15400.10270.11880.17280.05330.06970.13640.05510.1078
SV ↓PD-LTS0.18900.10650.14700.18570.05840.06370.12670.04320.1150
SV ↓PRISM0.18030.02540.09450.17540.02740.02850.12260.03310.0859
IPR (%) ↓Noisy12.999.585.2224.9230.7323.4723.0021.0818.87
IPR (%) ↓PGD1.323.540.390.3729.7719.2518.7820.2411.71
IPR (%) ↓PD-LTS6.786.681.6113.6426.8422.1714.9620.1514.10
IPR (%) ↓PRISM1.260.050.120.091.910.190.671.160.68
AuthorsOpenReview ↗

Response to Question 2

We appreciate the reviewer's careful examination of the bounded reliability calibration and its potential limitations. Bounded calibration is a deliberate design choice in GALR. It combines a local-spacing prior with structural descriptors, latent features, and object-level context through bounded logit corrections, allowing reliability estimation to remain grounded in geometric evidence while incorporating contextual adjustments. GALR performs initial geometric pre-repair, while GLFM further refines the latent representation.

To evaluate this constraint, we measured the proportion of initially inaccurate anchors affected by the calibration bound, obtaining 16.01%, 28.35%, and 31.64% under Low, Mid, and High corruption, respectively. GLFM further reduces the pre-decoding latent-anchor error from 14.970, 18.022, and 46.221 to 8.214, 4.482, and 9.697, corresponding to reductions of 45.13%, 75.13%, and 79.02%. These results indicate that although bounded calibration may limit some initial corrections, subsequent GLFM transport still provides substantial geometric refinement.

Regarding tightly clustered outliers, previous single-photon LiDAR studies indicate that background-induced false detections are generally stochastic and exhibit lower local spatial density, whereas genuine signal points tend to cluster along object surfaces (Wang et al., 2018; Goyal et al., 2025). Therefore, tightly clustered, high-density outliers are not a typical corruption pattern arising from the stochastic background-photon noise considered in SPORE.

For structures with limited local geometric support, GALR incorporates linearity, planarity, scattering, and object-level context for reliability estimation. GLFM further refines the geometry through global-local latent transport, allowing anchors to be updated even when their initial corrections are limited. This capability is illustrated by the chair example in Figure 3. Under High corruption, missing observations disrupt the continuity of thin structures, particularly the chair legs. PRISM effectively restores the continuity of the chair legs and supporting structures, providing qualitative evidence of its ability to recover geometry in regions with insufficient local observations.

Reviewer_Rq3LOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper makes two contributions: (1) SPORE, a physics-grounded simulated benchmark for single-photon point-cloud restoration, built from ShapeNet meshes with a ray-based SPAD acquisition model and Poisson photon statistics, spanning three corruption regimes. (2) PRISM, a hierarchical latent flow-matching framework that (i) estimates per-anchor reliability and selectively repairs unreliable local anchors while recording explicit token-wise repair states, and (ii) restores global and local latents via coupled velocity fields conditioned on those repair states.

Critical strengths
  1. Prior single-photon work targets multi-view reconstruction or scene modeling, while point-cloud denoising is almost universally evaluated using synthetic Gaussian/uniform perturbations. A physics-grounded, object-level benchmark with paired clean geometry data and controlled sensor-driven corruption is a genuinely useful resource.

  2. The scale is adequate, the three regimes jointly vary the degradation factors, and the paper validates that its corruption is more spatially heterogeneous than isotropic Gaussian noise.

Critical weaknesses
  1. The method does not win under mild degradation, so its value is confined to severe corruption. Under the Low-corruption regime, PRISM is clearly worse than PGD and PD-LTS on CD-L2 and P2M; it is best only on HD95 and ties SOR on [email protected]. The abstract’s claim of being “competitive under mild degradation” understates this limitation.

  2. The method is incremental relative to LION + flow matching. Hierarchical global/local latents, linear-interpolation flow matching, KNN Transformer, AdaLN, and FiLM are all existing components. The genuine novelty is reliability estimation, selective anchor repair, and persistent repair-state conditioning. The overall pipeline (two-stage training, frozen GALR during GLFM training, and a large number of hyperparameters) is complex, and the ablations cover only a few of these design choices.

  3. The method and the self-proposed benchmark appear in the same paper, with no independent validation. All strong quantitative conclusions rest on the authors’ own synthetic benchmark, and all nine baselines are adapted and retrained by the authors. The real-world experiment in Fig. 4 is a single scene with purely qualitative results, so the simulation-to-real transfer claim is weakly supported.

  4. Two-stage training may lead to accumulated error. GALR is frozen before GLFM training, so the repair states C never adapt during latent transport; the cost of this design choice relative to end-to-end training is not examined.

Questions for authors
  1. Please provide quantitative real-data results, not only a single qualitative figure. Otherwise, the transfer claim is not established.

  2. What causes the underperformance in the Low regime relative to PGD/PD-LTS? Is GALR’s reliability gating over-correcting observations that should be preserved, or does flow matching introduce bias near clean inputs? A mechanistic analysis of the Low regime would clarify this.

  3. How does joint end-to-end training compare with the current two-stage scheme in terms of performance and stability?

3 comments
AuthorsOpenReview ↗

Response to Question 1

Thank you for this comment. We tested PRISM, PGD, and PD-LTS on eight real, experimentally acquired single-photon point-cloud samples. Since clean ground truth has not yet been obtained, we used two no-reference geometric descriptors, local surface variation (SV) and isolated-point ratio (IPR).

After removing exact duplicate points, SV was calculated using each point’s 20 nearest neighbors, excluding the query point itself. It is the ratio of the smallest neighborhood covariance eigenvalue to the sum of all eigenvalues, averaged over all points. Lower SV indicates lower local nonplanarity. IPR measures the percentage of points with fewer than three other neighbors within a specified radius. This radius was set to three times the median nearest-neighbor distance in the deduplicated noisy input and was shared by all methods within each sample. Lower IPR indicates fewer isolated points.

The following table reports the unweighted means across the eight samples. PRISM achieved the lowest IPR on 8/8 samples and the lowest SV on 6/8 samples. These results indicate that PRISM continues to outperform PGD and PD-LTS overall in terms of SV and IPR on the evaluated real single-photon point clouds, providing additional no-reference evidence of its denoising effectiveness in this setting. Detailed visualizations and analysis, together with the quantitative comparison, will be provided in Section 5.3, “REAL-WORLD RESULTS,” of the revised manuscript.

MetricMethodSample 01Sample 02Sample 03Sample 04Sample 05Sample 06Sample 07Sample 08Mean
SV ↓Noisy0.19690.11360.14860.19440.06750.07190.13580.06100.1237
SV ↓PGD0.15400.10270.11880.17280.05330.06970.13640.05510.1078
SV ↓PD-LTS0.18900.10650.14700.18570.05840.06370.12670.04320.1150
SV ↓PRISM0.18030.02540.09450.17540.02740.02850.12260.03310.0859
IPR (%) ↓Noisy12.999.585.2224.9230.7323.4723.0021.0818.87
IPR (%) ↓PGD1.323.540.390.3729.7719.2518.7820.2411.71
IPR (%) ↓PD-LTS6.786.681.6113.6426.8422.1714.9620.1514.10
IPR (%) ↓PRISM1.260.050.120.091.910.190.671.160.68
AuthorsOpenReview ↗

Response to Question 2 and Question 3

Q2: We thank the reviewer for the careful observations regarding PRISM’s performance under Low corruption. PRISM was motivated by the severe, structured, and spatially heterogeneous degradation that can arise in real single-photon LiDAR acquisition. Limited photon counts, background illumination, and variations in surface reflectance can cause spurious detections, missing returns, and nonuniform point distributions (Tachella et al., 2019; Chen et al., 2020; Goyal et al., 2025). These effects extend beyond coordinate perturbations to include structural gaps and numerous unreliable points, motivating restoration that addresses both geometric accuracy and surface coverage.

Under Low corruption, PGD and PD-LTS achieve lower CD-L2 and P2M values, whereas PRISM achieves higher [email protected] and lower HD95, indicating a trade-off across geometric evaluation metrics. GALR provides conservative pre-repair through bounded calibration and continuous repair weights that limit unnecessary initial corrections. GLFM subsequently refines the representation using global shape information, reliability-aware neighborhood interactions, and repair-state conditioning. A small initial repair weight does not prevent an anchor from being updated during GLFM transport.

The Low-regime ablations support the positive contributions of this two-stage design. Removing GALR and its conditioning pathways increases CD-L2 and P2M from 1.164 and 1.240 in the full model to 3.511 and 3.190, respectively. From GALR-only to the full model, CD-L2 decreases from 1.183 to 1.164, P2M from 1.256 to 1.240, and HD95 from 0.0393 to 0.0279, while [email protected] increases from 0.1970 to 0.1997. Thus, subsequent GLFM transport yields a net improvement across all four metrics. Furthermore, removing persistent repair-state conditioning while retaining the same pre-repaired initialization increases HD95 to 0.0392, supporting the contribution of retained repair information to controlling large residual errors.

These results support the complementary roles of conservative pre-repair and subsequent latent refinement. Under Low corruption, PRISM’s advantages are most apparent in threshold-based correspondence and the control of large residual errors. Under High corruption, PRISM achieves the best performance among all compared methods across all four metrics, consistent with its design objective of restoring severely and structurally corrupted geometry.

Q3: We thank the reviewer for raising this important question. In our preliminary experiments, we explored joint end-to-end optimization of GALR and GLFM but observed substantial fluctuations in training dynamics and unstable gradients. These observations motivated our adoption of a two-stage training strategy. The instability may arise from the strong optimization dependencies between the two modules. GALR produces both pre-repaired local latents and reliability-aware repair states that condition GLFM. During joint optimization, updates to GALR continuously alter the latent initializations, conditioning signals, and local correspondences used to construct GLFM's training targets. Such non-stationarity may complicate velocity-field learning. Furthermore, GALR's geometric repair and reliability objectives differ from GLFM's latent restoration objectives, potentially introducing competing gradient signals during joint backpropagation.

We therefore optimize GALR first to establish geometric pre-repair and reliability estimation, then freeze its parameters while training GLFM. This allows the global and local velocity fields to learn from fixed pre-repaired initializations and consistent repair-state conditions, reducing the sources of non-stationarity encountered during joint optimization.

AuthorsOpenReview ↗

Additional Real-World Quantitative Evaluation

We collected single-view reference point clouds for five real objects (Samples 01 and 05–08) using a RealSense D457. We evaluated Noisy, PD-LTS, PGD, and PRISM using five metrics: CD-L2, [email protected], [email protected], [email protected], and HD95. PRISM achieves the best results in all 25 sample-metric comparisons across the five objects.

The quantitative results and corresponding qualitative visualizations will be added to Section 5.3 of the revised manuscript.

Quantitative results on real-world point clouds:

SampleMethodCD-L2 ↓[email protected] (%) ↑[email protected] (%) ↑[email protected] (%) ↑HD95 ↓
01Noisy0.3421.9711.9744.651.548
01PD-LTS0.2081.8211.8343.220.466
01PGD0.0142.5113.2145.450.170
01PRISM0.0115.1714.1846.290.103
02Noisy2.8963.1816.4250.785.047
02PD-LTS3.0003.9319.1357.505.025
02PGD2.8323.3818.4751.834.982
02PRISM0.00811.5334.8677.460.190
03Noisy4.4780.564.4926.964.994
03PD-LTS4.4381.155.9328.914.940
03PGD3.9521.015.8529.324.916
03PRISM0.0256.2615.5550.680.270
04Noisy0.5631.007.0534.702.209
04PD-LTS0.5141.257.9439.152.166
04PGD0.4261.018.0838.961.857
04PRISM0.0186.2213.3848.600.215
05Noisy1.5240.939.4340.863.795
05PD-LTS1.6161.379.9842.583.993
05PGD0.8471.199.5241.752.574
05PRISM0.0148.8214.4553.070.182
Reviewer_dmErOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a point-cloud restoration framework for degraded single-photon measurements. The method combines Geometry-Aware Latent Repair (GALR), which estimates local geometric reliability and selectively corrects corrupted anchors, with Global-Local Flow Matching (GLFM), which restores hierarchical point-cloud representations using persistent repair-state conditioning. The authors also introduce SPORE, a simulated single-photon point-cloud benchmark constructed from ShapeNet objects using ray-based acquisition and photon-counting models.

Experiments across three corruption levels show that PRISM achieves the best performance on three of four metrics under Mid corruption and all four metrics under High corruption. Ablation studies further demonstrate the contributions of geometry-aware repair, repair-state conditioning, and reliability-aware feature processing.

Critical strengths
  • Well-motivated reliability-aware restoration framework. The proposed method explicitly accounts for spatially varying corruption rather than treating all observed points equally. Combining selective local repair with global-local latent restoration is a reasonable approach to recovering geometry under severe degradation.

  • Strong results under severe corruption with meaningful ablations. PRISM achieves consistent improvements over existing point-cloud restoration methods under High corruption, particularly in F-score and HD95. The ablations separately investigate repair-state conditioning, reliability-aware neighborhoods, and point-wise FiLM, providing useful evidence for the effectiveness of these components.

Critical weaknesses
  • Limited evidence of real-world generalization. The evaluation is conducted almost entirely on the proposed synthetic SPORE benchmark. Although the simulator incorporates photon statistics and acquisition-related corruption, it is unclear how closely these simulated measurements match real single-photon sensor characteristics. In particular, the benchmark uses ten simulated viewpoints per object and fixed-size point-cloud preprocessing, which may not reflect practical acquisition conditions. The real-world evaluation provides only a qualitative demonstration without ground-truth comparisons or quantitative metrics. This limits the evidence supporting practical applicability.

  • Insufficient evaluation of robustness and computational trade-offs. The method combines several components, including reliability estimation, anchor correction, hierarchical latent representations, and coupled flow-matching networks. Although component ablations are provided, the paper does not sufficiently evaluate whether this complexity is justified relative to simpler restoration approaches in terms of computational cost and inference latency. Furthermore, performance is primarily evaluated under the same corruption model and regimes used during training. Generalization to unseen corruption distributions, acquisition configurations, and sensor characteristics remains unclear.

Questions for authors

Can the authors provide quantitative evaluations on real single-photon measurements, ideally with reference geometry or comparisons against existing restoration methods? This would help establish whether improvements on SPORE transfer to actual acquisition conditions.

How robust is PRISM to corruption distributions or acquisition parameters not encountered during training, such as different photon-count rates, background noise levels, viewpoint configurations, or sensor noise characteristics?

Can the authors report inference latency and computational complexity relative to the strongest competing methods? Given the multi-stage architecture, it would be useful to understand the practical cost of the reported accuracy improvements.

3 comments
AuthorsOpenReview ↗

Response to Question 1 and Question 2

Q1: Thank you for this comment. We tested PRISM, PGD, and PD-LTS on eight real, experimentally acquired single-photon point-cloud samples. Since clean ground truth has not yet been obtained, we used two no-reference geometric descriptors, local surface variation (SV) and isolated-point ratio (IPR).

After removing exact duplicate points, SV was calculated using each point’s 20 nearest neighbors, excluding the query point itself. It is the ratio of the smallest neighborhood covariance eigenvalue to the sum of all eigenvalues, averaged over all points. Lower SV indicates lower local nonplanarity. IPR measures the percentage of points with fewer than three other neighbors within a specified radius. This radius was set to three times the median nearest-neighbor distance in the deduplicated noisy input and was shared by all methods within each sample. Lower IPR indicates fewer isolated points.

The following table reports the unweighted means across the eight samples. PRISM achieved the lowest IPR on 8/8 samples and the lowest SV on 6/8 samples. These results indicate that PRISM continues to outperform PGD and PD-LTS overall in terms of SV and IPR on the evaluated real single-photon point clouds, providing additional no-reference evidence of its denoising effectiveness in this setting. Detailed visualizations and analysis, together with the quantitative comparison, will be provided in Section 5.3, “REAL-WORLD RESULTS,” of the revised manuscript.

MetricMethodSample 01Sample 02Sample 03Sample 04Sample 05Sample 06Sample 07Sample 08Mean
SV ↓Noisy0.19690.11360.14860.19440.06750.07190.13580.06100.1237
SV ↓PGD0.15400.10270.11880.17280.05330.06970.13640.05510.1078
SV ↓PD-LTS0.18900.10650.14700.18570.05840.06370.12670.04320.1150
SV ↓PRISM0.18030.02540.09450.17540.02740.02850.12260.03310.0859
IPR (%) ↓Noisy12.999.585.2224.9230.7323.4723.0021.0818.87
IPR (%) ↓PGD1.323.540.390.3729.7719.2518.7820.2411.71
IPR (%) ↓PD-LTS6.786.681.6113.6426.8422.1714.9620.1514.10
IPR (%) ↓PRISM1.260.050.120.091.910.190.671.160.68

Q2: We thank the reviewer for raising this important question. To provide a systematic evaluation, we have added a dedicated section, “Robustness to Simulator and Acquisition Shifts,” to the revised Appendix. We conduct 20 Mid-anchored experiments, comprising one nominal reference, 12 single-factor perturbations, three combined corruption settings, and four acquisition-geometry variations. These experiments cover changes in photon emission, background noise, dropout, timing and angular jitter, spurious detections, viewpoint configurations, and field of view. All evaluations use the same trained PRISM model without additional fine-tuning.

The results demonstrate PRISM's consistent restoration capability across the tested parameter shifts. Across all 19 shifted conditions, PRISM improves all four evaluation metrics relative to the corresponding noisy inputs, reducing CD-L2 by 93.4–97.9% and P2M by 92.9–97.9%. For example, increasing the background rate from 0.15 to 0.60 raises the noisy CD-L2 from 32.813 to 66.762, while PRISM maintains a low CD-L2 of 1.421. Even under the combined low-photon and high-dropout condition, PRISM reduces CD-L2 from 180.318 to 8.318 and HD95 from 0.8230 to 0.0387. With only three acquisition views, it further reduces CD-L2 from 38.902 to 2.571 while improving [email protected] from 0.0833 to 0.2421.

These results provide quantitative evidence that PRISM maintains effective geometric restoration across diverse simulator configurations, including substantial changes in photon statistics, sensor noise, and acquisition geometry. The complete parameter settings, quantitative comparisons, and detailed analyses are now included in the updated Appendix.

AuthorsOpenReview ↗

Response to Question 3

We thank the reviewer for highlighting the importance of computational efficiency, particularly given the multi-stage architecture of PRISM. To evaluate computational efficiency, we benchmarked PRISM, PGD, and PD-LTS under the same experimental configurations as those used in our comparison study, using a single NVIDIA RTX PRO 6000 GPU with a batch size of 1. The reported inference latency covers the complete inference pipeline, including all network stages. GPU warm-up was performed before measurement, and GPU synchronization was used to ensure accurate latency measurements. We report both end-to-end inference latency and computational complexity (GFLOPs), as summarized below.

MethodInference Latency (ms) ↓FLOPs (GFLOPs) ↓
PRISM207.91310.01
PGD419.45181.65
PD-LTS189.15132.25

PRISM achieves an inference latency of 207.91 ms, making it approximately 2.02 times faster than PGD (419.45 ms), despite requiring more floating-point operations (310.01 vs. 181.65 GFLOPs). Compared with PD-LTS, PRISM incurs approximately 9.9% higher inference latency (207.91 vs. 189.15 ms) and 2.34 times the computational cost in GFLOPs.

These results indicate that PRISM's improved restoration performance comes with increased arithmetic complexity. Nevertheless, its measured inference latency remains substantially lower than that of PGD and relatively close to that of PD-LTS under the evaluated hardware and experimental settings.

AuthorsOpenReview ↗

Additional Real-World Quantitative Evaluation

We collected single-view reference point clouds for five real objects (Samples 01 and 05–08) using a RealSense D457. We evaluated Noisy, PD-LTS, PGD, and PRISM using five metrics: CD-L2, [email protected], [email protected], [email protected], and HD95. PRISM achieves the best results in all 25 sample-metric comparisons across the five objects.

The quantitative results and corresponding qualitative visualizations will be added to Section 5.3 of the revised manuscript.

Quantitative results on real-world point clouds:

SampleMethodCD-L2 ↓[email protected] (%) ↑[email protected] (%) ↑[email protected] (%) ↑HD95 ↓
01Noisy0.3421.9711.9744.651.548
01PD-LTS0.2081.8211.8343.220.466
01PGD0.0142.5113.2145.450.170
01PRISM0.0115.1714.1846.290.103
02Noisy2.8963.1816.4250.785.047
02PD-LTS3.0003.9319.1357.505.025
02PGD2.8323.3818.4751.834.982
02PRISM0.00811.5334.8677.460.190
03Noisy4.4780.564.4926.964.994
03PD-LTS4.4381.155.9328.914.940
03PGD3.9521.015.8529.324.916
03PRISM0.0256.2615.5550.680.270
04Noisy0.5631.007.0534.702.209
04PD-LTS0.5141.257.9439.152.166
04PGD0.4261.018.0838.961.857
04PRISM0.0186.2213.3848.600.215
05Noisy1.5240.939.4340.863.795
05PD-LTS1.6161.379.9842.583.993
05PGD0.8471.199.5241.752.574
05PRISM0.0148.8214.4553.070.182

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers

We sincerely thank the reviewers for their constructive comments. In response, we have revised the manuscript and highlighted all additions and modifications in blue for clarity. The main updates are summarized below:

Section 5.3 (Real-World Results): We expanded the real-world evaluation with eight single-photon point clouds acquired using our SPAD-based system. We added reference-free quantitative comparisons using SV and IPR, together with qualitative results on three representative objects.

Appendix (Robustness to Simulator and Acquisition Shifts): We added robustness experiments covering 19 simulator and acquisition shifts. All experiments use the same frozen PRISM checkpoint and test-object set, without retraining or fine-tuning.

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!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.