acceptodds
Under review as a conference paper at ICLR 2027

BeamScene: A Physically Consistent Framework with Beam-Footprint Geometry for 3D LiDAR Scene Completion

Abstract

Autonomous vehicles depend on LiDAR for metrically accurate 3D perception, yet a single scan is sparse, and occluded regions are left almost empty. LiDAR scene completion recovers the missing geometry, but diffusion-based completion requires many network evaluations per scan, and recent single-pass methods place candidate points, choose feature scales, and supervise visibility with hand-set rules that are unrelated to how the sensor samples the scene. We present BeamScene, a single-pass framework that derives each of these decisions from one quantity, the beam footprint of every return, obtained from its range and the angular resolution of the sensor. The Beam-Footprint-Aware Candidate Constructor (BFCC) allocates an exact, bounded candidate budget in proportion to the footprint, spreads candidates in the ray-orthogonal tangent plane, and keeps every measured return as an immutable anchor. The Footprint-Aligned Shared-Scale Encoder (FASSE) shares sparse kernels across eight voxel resolutions and both input branches, and weights the resolutions with a parameter-free footprint-to-grid gate. The Frustum-Ray Visibility Objective (FRVO) adds frustum-local entropic transport and differentiable first-return and free-space terms during training only. On SemanticKITTI, BeamScene raises voxel IoU at and from and (LiNeXt) to and , in a regime where Chamfer distance has largely saturated; its refined variant is the best of all compared methods on all six metrics of both SemanticKITTI and KITTI-360. The coarse model has M parameters, fewer than LiNeXt, and runs in per scan, – faster than diffusion-based completion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LiDAR Scene Completion, Beam-Footprint Geometry, Efficient Single-Pass 3D Completion, Visibility-Aware Supervision

Reviews

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

This paper studies dense scene completion from a single sparse LiDAR scan and proposes BeamScene, a single-pass framework that uses an approximate beam footprint, derived from range and sensor angular resolution, to coordinate candidate generation, feature-scale selection, and geometric supervision. The framework contains three main components. BFCC allocates a fixed candidate budget according to footprint area, expands candidates in ray-orthogonal planes, and preserves input observations as immutable anchors. FASSE shares sparse convolutional kernels across voxel scales and input branches, combining features through footprint-dependent scale weights. FRVO introduces training-only supervision based on local optimal transport, first-return consistency, and a free-space penalty. An observation-preserving coarse-to-fine decoder produces the completed point clouds. The paper evaluates the method on SemanticKITTI and tests transfer to KITTI-360 without fine-tuning. The coarse model’s main gains over LiNeXt occur in fine-resolution voxel IoU, alongside a 43.7% reduction in parameters. The refined model achieves the best reported results across all six metrics among the compared methods on both datasets. Component ablations, internal design ablations, and efficiency comparisons provide further evaluation.

Critical strengths
  1. Meaningful completion improvements with substantially fewer parameters. On SemanticKITTI, the coarse model improves [email protected] from 19.45 to 24.91 and [email protected] from 6.30 to 8.74 relative to LiNeXt, while reducing parameters from 1.99M to 1.12M. Evaluation on KITTI-360 without fine-tuning shows improvements in the same direction. This establishes practical value through improved geometric recovery alongside parameter reduction. The latency reduction relative to LiNeXt is modest; the principal efficiency benefits are parameter savings and substantial acceleration relative to diffusion-based methods.

  2. The architecture incorporates useful engineering properties. Exact preservation of input observations, control over output cardinality, and restriction of additional visibility supervision to training are meaningful design choices. They make the relationship between measured and generated geometry explicit while maintaining a straightforward inference pipeline. The appendix also helps clarify which properties follow directly from construction and where the method’s guarantees end.

Critical weaknesses
  1. The central physical-consistency claim lacks direct validation, and the FRVO surrogate has potentially consequential limitations. Section 3.5 and Table 4 define “rays” using 1∘×1∘1^\circ\times1^\circ angular bins, which are substantially coarser than the nominal angular resolution used by BFCC. A single bin can contain both foreground and background surfaces, so its minimum range does not directly represent the first return along every direction within that bin. Furthermore, the reference geometry comes from aggregated scans, and the loss involves point subsampling. It remains unclear how these operations preserve reliable free-space evidence from the input viewpoint. The free-space term in Eq. (6) constrains a softmax-weighted average range. Consequently, an early point can be diluted by sufficiently many slightly farther points. For example, using the paper’s temperature, a bin containing one point at 9.8m and 127 points at 10m has a soft first return of approximately 9.989m. With a reference range of 10m and a tolerance of 0.1m, the one-sided free-space penalty is zero despite the point appearing 0.2m before the reference. This is an illustrative example derived from the formula, not an observed experimental result, but it demonstrates a possible failure of the surrogate. The authors acknowledge that inference-time consistency is not guaranteed. My concern is the absence of direct validation: CD, JSD, and IoU do not establish whether actual free-space violations or first-return errors decrease. The experiments support improved completion metrics, but the central physical interpretation remains insufficiently substantiated.

  2. The internal ablations do not sufficiently establish the necessity or specific benefits of the unified footprint prior. In Table 2(c), replacing footprint routing with uniform weights improves IoU at all three voxel resolutions, including an increase in [email protected] from 8.74 to 9.44, and also improves JSD 3D. In Table 2(d), CD + FT achieves an [email protected] of 9.52, exceeding the full objective. The authors appropriately acknowledge these trade-offs. Nevertheless, because fine-resolution occupancy is a central motivation and headline result, the selection of the final configuration needs stronger justification. Additionally, under uniform angular resolution, the common factor ΔθΔϕ\Delta\theta\Delta\phi cancels when the clipped footprint areas are normalized into allocation weights. Candidate-count allocation therefore reduces to bounded range-squared weighting. Angular resolution still affects expansion radius and scale routing, but candidate-count allocation itself has limited sensor specificity. Together with the absence of repeated-run variability, these results leave uncertainty about whether the full physical design offers stable advantages over simpler alternatives.

Questions for authors
  1. Validity of visibility supervision: How do you handle foreground/background mixing within an angular bin and subsampling that omits the nearest point? Can you report free-space violation rates and first-return errors evaluated using the input’s actual observed rays, comparing LiNeXt, the full model, and the model without the Ray term? This would directly affect my assessment of the physical-consistency contribution.

  2. Configuration selection and stability: Given that uniform scale weighting and CD + FT outperform the full configuration on several important metrics, what criterion determined the final design? Are the key differences in Table 2 stable across repeated training runs, using consistent training budgets and model-selection procedures?

  3. Specific role of footprint modeling: Do you agree that uniform angular resolution cancels from the candidate-count allocation weights? After adaptive voxel downsampling to 18,000 input points, how does the nominal sensor footprint correspond to the actual input sampling spacing? Which existing results distinguish the benefits of footprint modeling from those of a generic range-adaptive strategy?

4 comments
AuthorsOpenReview ↗

Response to Reviewer khLs (1/4)

We thank the reviewer for a careful review that engages with our equations directly. Several observations are correct as stated, and we address them using the formulation and results in the paper, together with a new direct visibility evaluation. Tables are numbered consistently across our responses to all reviewers. We have uploaded a revised PDF in which all changes are marked in blue; the new tables keep the numbering of this response (Tables R1–R7, with R2 and R5 in the main text and the others in Appendix C). Items that need new experiments or renderings (three seeds, the cumulative ablation from LiNeXt, the new insets in Fig. 5, and cross-sensor evaluation) will be added in the final version.

W1 / Q1 — Validity of the visibility supervision

(a) Mixed foreground/background bins make the free-space penalty conservative. In Eq. (6), the reference first return of a bin is rb⋆=min⁡y∈Yb∥y∥r_b^\star=\min_{y\in Y_b}\|y\|, the nearest surface anywhere in the bin. Every direction in the bin passes through [0,rb⋆)[0,r_b^\star) without a return, whichever surface it eventually hits. The free-space penalty acts only on predictions in [0,rb⋆−δf)[0,r_b^\star-\delta_f), so it never penalises a prediction at a range where some ray in the bin was blocked. Mixing has one effect: in a bin containing a foreground surface, the free space in front of the background directions, between the two ranges, is left unconstrained. In other words, the coarse bin leads to missed penalties, not wrong ones. This holds for the full reference; subsampling is discussed in (c).

(b) Relation to the input viewpoint after aggregation. The reference is built by aggregating neighbouring scans and removing dynamic objects (Sec. 4.1). Following LiDiff, moving objects are removed from the input scan with the same labels, so every input return lies on a static surface that the reference contains. For the full reference, in any bin that contains input returns, rb⋆r_b^\star is therefore at or before each of those returns. Each input ray certifies [0,rin)⊇[0,rb⋆)[0,r_{\text{in}})\supseteq[0,r_b^\star) as empty. The penalised region is therefore a subset of the free space observed from the input viewpoint, up to registration error, which δf\delta_f absorbs. Aggregation can only move rb⋆r_b^\star closer, which shrinks the penalised region and never enlarges it. In bins with no input return, the free-space penalty enforces only consistency with the supervision target (“no reference surface in front of rb⋆r_b^\star”). The revised Sec. 3.5 states this distinction.

Why 1∘×1∘1^\circ\times1^\circ rather than the nominal resolution. The ray term (first-return term plus free-space penalty, Eq. 6) uses only bins that are non-empty in both sets, after each set is capped at 16,384 points (App. A). At the nominal angular resolution of 0.20∘×0.42∘0.20^\circ\times0.42^\circ used by BFCC there are about 1,800×64≈1151{,}800\times64\approx115k directions, so on average 0.14 points per direction, and almost no direction would be occupied in both sets. A 1∘×1∘1^\circ\times1^\circ bin, about 5×2.45\times2.4 beam spacings at nominal resolution, is roughly the finest grid at which common rays exist in useful numbers. As App. A states, the FRVO partitions are intentionally independent of the BFCC discretisation.

(c) Subsampling. The ray term caps both the predicted and the reference sets at the same budget before binning, so omitting the nearest point affects both first-return estimates alike in expectation, but not in a given bin. With about 1.7 points per bin on average (see (d)), the subsampled reference minimum can lie behind the nearest reference surface. Because the free-space penalty is one-sided, such fluctuations do not average out: they can penalise a prediction that lies on the nearest surface. Subsampling can therefore introduce wrong penalties, and the conservative reading in (a) and (b) holds for the full reference only. The revised Sec. 3.5 states this. Keeping the nearest reference point of each bin when subsampling would remove the effect at no extra cost. The net effect of the objective on actual violations is measured directly in (e).

AuthorsOpenReview ↗

Response to Reviewer khLs (2/4)

W1 / Q1 — Validity of the visibility supervision (cont.)

(d) The dilution example is correct, and its extent is bounded. We confirm the arithmetic: r^b≈9.989\hat r_b\approx 9.989 m, and the free-space penalty is zero. From Eq. (6), a single point at depth dd in front of rb⋆r_b^\star escapes the penalty only if at least

n  ≥  (dδf−1)ed/τrn \;\ge\; \left(\frac{d}{\delta_f}-1\right)e^{d/\tau_r}

other points in the same ray lie at rb⋆r_b^\star. Points placed farther behind dilute less, so this is the minimum. With τr=δf=0.1\tau_r=\delta_f=0.1 m (Table 4), this requires 8 points for d=0.2d=0.2 m, 41 for 0.3 m and 164 for 0.4 m. Because the ray term keeps at most 128 points per ray (App. A), your example of 1+1271+127 points sits exactly at the cap. A violation deeper than about 0.38 m cannot be masked in any ray. For scale, 16,384 points spread over the roughly 360∘×27∘360^\circ\times27^\circ field of view give about 1.7 points per 1∘×1∘1^\circ\times1^\circ bin on average, far below the 8 points needed to mask even a 0.2 m violation. A masked shallow violation is also still penalised by the other two terms: the point pays its distance to the nearest reference point, at most 0.2 m, in the Chamfer term, and pays the radial cost λrad∣ra−rb∣/ℓc\lambda_{\text{rad}}|r_a-r_b|/\ell_c of frustum-local transport, Eq. (5). Dilution therefore lowers the sensitivity of one term to shallow violations in dense rays; it does not create a blind spot in the objective. The direct evaluation in (e) measures the net effect on actual violations. The revised PDF adds your example and this bound to Appendix C (referenced from Sec. 3.5). It also notes that a per-point hinge ∑i[rb⋆−δf−rb,i]+\sum_i[r_b^\star-\delta_f-r_{b,i}]_+ removes dilution entirely.

(e) Direct validation on the input's actual rays. Following your suggestion, we evaluate on the ∼102{\sim}102k rays of each raw scan that are not among the 18k input points, i.e., observations the model never sees. A ray has a free-space violation if a predicted point lies within its beam cone (0.001 rad), at a range below rraw−δr_{\text{raw}}-\delta, and farther than δs=0.1\delta_s=0.1 m from every reference point. The last condition excludes points on real surfaces, which at grazing incidence span a long range interval within the cone. A ray has a missing first return if no predicted point lies within δs\delta_s of its measured return. Both rates are computed over all evaluated rays. Predicting fewer points would lower the violation rate but raise the missing rate, so the two are read together. All models are coarse (without refinement, as in Table 2), with comparable output sizes (∼177{\sim}177k–180k points). Statistics are pooled over all frames of seq. 08. As a sanity check, the raw scan itself scores 0 on both metrics.

Table R5. Visibility on unseen raw-scan rays, SemanticKITTI seq. 08 (coarse models, all frames).

MethodFree-space violation ↓\downarrow (%) @0.1 m / 0.2 m / 0.5 mMissing first return ↓\downarrow (%)
LiNeXt0.63 / 0.60 / 0.5136.16
BeamScene, CD only0.59 / 0.57 / 0.4835.30
BeamScene, CD + FT (w/o ray term)0.54 / 0.53 / 0.4634.67
BeamScene (full)0.56 / 0.51 / 0.4334.39

Three observations:

  1. BeamScene reduces actual violations relative to LiNeXt by 11–16% at every threshold, and lowers the missing rate by 1.8 points. On the 18k input rays, BeamScene's missing rate is exactly 0 by Prop. 2, whereas LiNeXt displaces measured points.
  2. Adding the ray term lowers deep violations without increasing missing returns. Compared with CD + FT (w/o ray term), violations decrease at 0.2 m (0.53→0.51%0.53\to0.51\%) and 0.5 m (0.46→0.43%0.46\to0.43\%), and the missing rate also decreases slightly (34.67→34.39%34.67\to34.39\%), so the reduction does not come from predicting fewer points. These differences are 0.02–0.03 percentage points from a single training run. We read them as consistent with the design of Eq. (6), not as established, and include this evaluation in the three-seed runs (W2).
  3. The shallow band behaves as Eq. (6) implies. At 0.1 m, the full model is slightly higher (0.56 vs. 0.54%), i.e., violations shift from >0.5{>}0.5 m into the 0.1–0.2 m band. The free-space penalty is a squared hinge that starts at δf=0.1\delta_f=0.1 m, so it acts weakly on violations just beyond 0.1 m and strongly on deep ones.
AuthorsOpenReview ↗

Response to Reviewer khLs (3/4)

W1 / Q1 — Validity of the visibility supervision (cont.)

Under this strict definition, violations are rare (<1%{<}1\%) for all methods. We therefore read the comparison as consistent relative reductions, not as a large absolute gap. The revised PDF adds this table (Table R5) to Sec. 4.3, and states explicitly that “physically consistent” refers to how the pipeline is constructed (Sec. 1, L100–104); FRVO encourages the visibility property, as measured above, but does not guarantee it at inference (Sec. 4.5, App. B).

W2 / Q2 — How the configuration was selected, and stability

We agree that in Table 2(c)/(d) the full model is best only in CD and JSD BEV.

  1. Selection criterion. The full configuration was fixed from the footprint model before any evaluation, and sequence 08 was never used to select it. Rows (b)–(d) ablate this design afterwards; they are not candidates in a search. Our protocol (Sec. 4.1: train on 00–07 and 09–10, evaluate on 08) has no held-out validation split, so switching to the best-scoring row on sequence 08 would select on the test split, and we did not do so. The fixed configuration happens to be best in CD, the only metric reported by every baseline (Table 1), and in JSD BEV, the global layout metric, while the two variants are better in fine IoU.
  2. The comparison with LiNeXt does not depend on this choice. Both alternatives beat LiNeXt (19.45 / 6.30 at 0.2 / 0.1 m) by more than the full model does: uniform scale weighting (“w/o footprint routing” in Table 2c) reaches 25.51 / 9.44 and CD + FT reaches 25.70 / 9.52, against 24.91 / 8.74. Within Table 2, the largest fine-IoU effects come from BFCC and frustum-local transport: removing BFCC gives the lowest [email protected] m among the single-component removals (7.68), and frustum-local transport adds +1.53+1.53 over CD only. The footprint gate and the ray term trade some fine occupancy for CD and BEV layout, which Sec. 4.3 (L425–431) already describes as regularisation rather than a uniform gain. These are marginal effects, each measured with the other components present, and they do not account for the whole gain over LiNeXt at 0.2 m; the final version adds a cumulative ablation that starts from LiNeXt retrained in our implementation. The revised Sec. 4.3 states the selection criterion and presents uniform scale weighting and CD + FT as alternative operating points; the attribution will be stated in the contribution list of the final version.
  3. Stability. We agree that the submission lacks repeated runs. All variants follow the identical schedule in App. A (10 coarse epochs, the same optimiser and learning rate, batch size 16) and are evaluated at the final checkpoint, with no checkpoint selection on the test split. Variance estimates are therefore needed here. The largest component-level effects are clear: removing FRVO raises CD from 0.205 to 0.224, and removing BFCC lowers [email protected] m from 8.74 to 7.68. The smaller differences are not: removing FASSE changes CD by 0.003, the same size as the CD differences between the full model and the two variants (0.002 and 0.005), and the gaps between the full model and CD + FT in (e) are 0.02–0.03 percentage points. For fixed checkpoints, a paired block bootstrap over the scans of sequence 08 shows that the differences between the full model and these two variants hold consistently across scans (all 95% intervals exclude zero). This does not measure variation across training runs, so we do not claim that these smaller differences would persist under retraining, and we have worded the corresponding conclusions accordingly. For the final version, we will train the full model, uniform scale weighting, CD + FT and the three single-component removals of Table 2(a) with three seeds each, and report mean ±\pm std for Table 2 and for the visibility evaluation in (e).
AuthorsOpenReview ↗

Response to Reviewer khLs (4/4)

W3 / Q3 — What footprint modelling specifically contributes

We agree that angular resolution cancels. Under uniform resolution, Ai=ri2ΔθΔϕA_i=r_i^2\Delta\theta\Delta\phi, and the quantile bounds scale by the same constant. Hence wi=clip(ri2)/∑jclip(rj2)w_i=\mathrm{clip}(r_i^2)/\sum_j\mathrm{clip}(r_j^2), and candidate allocation is bounded range-squared weighting. The revised PDF states this in Sec. 3.2 and no longer describes the allocation as sensor-specific. Our claim is narrower: r2r^2 is not a tuned schedule but the area law of constant-angle sampling. The range-adaptive rule used in prior work, LiNeXt's range groups, is a much flatter step function (5 to 15 copies, a factor of 3) with group boundaries and counts set by hand, whereas BFCC assigns between 1 and nmax⁡=24n_{\max}=24 candidates per return.

Where angular resolution does matter. Above smin⁡s_{\min}, Eq. (1) gives log⁡si=log⁡ri+12log⁡(ΔθΔϕ)\log s_i=\log r_i+\tfrac12\log(\Delta\theta\Delta\phi). The angular resolution sets the offset of the range-to-scale mapping shared by the expansion radius in Eq. (2) and the gate in Eq. (3), i.e., which voxel size is preferred at which range. With the values in Table 4 (τs=0.70\tau_s=0.70, gk=0.01⋅2k−1g_k=0.01\cdot2^{k-1} m), the gate peaks at 0.04 m at 10 m and at 0.16 m at 40 m. Without the sensor model this offset would be another hyperparameter. For sensors with non-uniform or per-beam resolution, ΔθΔϕ\Delta\theta\Delta\phi varies across returns and no longer cancels from the allocation (Limitation 1).

Nominal footprint vs. the spacing of the 18,000 input points. By definition si2=ri2ΔθΔϕs_i^2=r_i^2\Delta\theta\Delta\phi is the approximate area that one beam covers at range rir_i under the nominal angular resolution (Eq. 1). For example, at 10 m the nominal beam spacing is 3.5 cm×7.3 cm3.5\,\text{cm}\times7.3\,\text{cm} and si=5.1s_i=5.1 cm. If the downsampling keeps a fraction ρ(r)\rho(r) of the returns near range rr, the actual area per retained return is Ai/ρ(ri)A_i/\rho(r_i). The footprint therefore matches the input spacing up to the retention ratio. Allocation is exactly area-proportional when ρ\rho does not depend on range, and only relative weights wiw_i enter it, so a uniform thinning has no effect. If voxel downsampling removes proportionally more near-range points, the effective area grows more slowly than r2r^2 at short range. The lower quantile clip (qℓ=0.05q_\ell=0.05) flattens exactly that end of the distribution. Routing is soft by design: with τs=0.70\tau_s=0.70, about 70% of the weight falls on two adjacent resolutions, so it does not require sis_i to match the spacing precisely. We will add this discussion, and the correction Ai/ρ(ri)A_i/\rho(r_i) as a direct extension, in the final version.

Which existing results separate footprint modelling from a generic range-adaptive rule.

  • Allocation. Table 2(b), “distance-grouped allocation”, isolates exactly this comparison: the tangent-plane expansion, encoder and loss are kept, and footprint allocation is replaced by LiNeXt's manually defined range groups (range quartiles with 5/8/12/15 copies). The r2r^2 law is better on all six metrics, mainly in global layout (CD 0.211→0.2050.211\to0.205, JSD BEV 0.341→0.3320.341\to0.332) and only marginally at [email protected] m (8.68→8.748.68\to8.74). Within BFCC, the larger fine-IoU effect comes from the tangent-plane expansion (isotropic expansion: 8.42).
  • Routing. Our ablation compares the footprint gate only with uniform scale weighting, which do not depend on range. It does not separate the gate from a range-based gate with a hand-tuned offset, and we do not claim it does. We will describe the gate as range-adaptive routing whose offset comes from the sensor model instead of tuning.

Summary of revisions

Already in the revised PDF (marked in blue).

  1. Sec. 3.5 and Appendix C: the conservative reading of FRVO for the full reference, the effect of point subsampling, the justification for 1∘×1∘1^\circ\times1^\circ bins, and your example with the masking bound.
  2. Sec. 4.3: the input-ray visibility evaluation in (e) (Table R5), with its protocol in Appendix C.
  3. Sec. 4.3: an explicit selection criterion, with uniform scale weighting and CD + FT as alternative operating points; Appendix C: paired confidence intervals (Table R7).
  4. Sec. 3.2: the reduction of allocation to clipped r2r^2 weighting under uniform resolution.

We thank the reviewer again; the questions have clarified the scope of our physical-consistency claim, and we are happy to discuss further.

Reviewer_kQbfOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BeamScene, a single-pass framework for completing sparse LiDAR scenes. It estimates an approximate beam-footprint scale from range and sensor angular resolution and uses this scale to guide candidate generation and multiscale feature fusion. BFCC allocates candidate points according to footprint area, expands them in ray-orthogonal planes, and preserves the input measurements. FASSE shares convolutional kernels across voxel resolutions and input branches, with footprint-dependent feature weighting. FRVO augments Chamfer supervision with frustum-local optimal transport, first-return consistency, and a free-space penalty. Experiments on SemanticKITTI and KITTI-360 without fine-tuning demonstrate competitive completion performance with substantially fewer parameters than the strongest single-pass baseline.

Critical strengths

1.BeamScene achieves meaningful improvements while substantially reducing the parameter count. Compared with LiNeXt, the coarse model reduces parameters from 1.99 M to 1.12 M while improving all six reported metrics on SemanticKITTI. In particular, voxel IoU at 0.2 m and 0.1 m increases from 19.45 and 6.30 to 24.91 and 8.74, respectively. Evaluation on KITTI-360 without fine-tuning provides additional evidence of cross-dataset generalization. 2.The framework has a coherent geometric motivation and practical design choices. Candidate allocation, scale fusion, and observation preservation are connected through the sensor sampling model. The fixed candidate budget and exact retention of input measurements provide predictable output size and observation fidelity, while shared multiscale kernels substantially reduce parameters. Both component-level and internal ablations provide a useful basis for examining these design choices.

Critical weaknesses

1.Table 2 does not fully establish why the complete FASSE and FRVO configurations should be preferred over simpler alternatives. Uniform scale weighting improves 0.1 m IoU from 8.74 to 9.44, while CD plus frustum-local transport achieves 9.52. Although the full model achieves better CD and BEV distribution metrics, a deeper explanation of these trade-offs would help justify the final configurations. 2.The method introduces numerous hyperparameters across candidate construction, scale routing, and geometric supervision. Their selection procedure, search ranges, and sensitivity are insufficiently documented, making it difficult to assess how much performance depends on careful tuning. The cross-dataset results are encouraging, but sensitivity analysis and clarification of the data splits used for tuning would strengthen confidence in robustness. 3.Improvements in overall IoU do not clearly distinguish densification of observed surfaces from recovery of unobserved geometry. Results broken down by distance or visibility would better support the paper’s interpretation that the method improves completion in distant and occluded regions. 4.Some gains over LiNeXt† are relatively small: on SemanticKITTI, CD decreases from 0.149 to 0.144, and 0.2 m IoU increases from 31.25 to 31.93. Repeated runs or appropriate uncertainty estimates would help establish the stability of these improvements, particularly for closely performing ablation configurations.

Questions for authors

1.What criteria motivated the selection of the complete FASSE and FRVO configurations, given that simpler variants perform better on several metrics? Can further analysis identify the scenes or geometric properties for which the complete configurations are preferable? 2.How were the main hyperparameters selected, and which data splits were used for tuning? Was a common configuration used across evaluations, and which parameters have the greatest influence on performance? 3.Can the authors provide results across distance ranges and observed versus unobserved regions to clarify the respective contributions of surface densification and occluded-region recovery? 4.Can repeated-run results or appropriate uncertainty estimates be provided for the key comparisons with LiNeXt† and for closely performing ablation configurations?

5 comments
AuthorsOpenReview ↗

Response to Reviewer kQbf (1/5)

We thank the reviewer for the careful reading and for recognising the efficiency, the cross-dataset results, and the geometric motivation of BeamScene. We address each point below; We have uploaded a revised PDF in which all changes are marked in blue; the new tables keep the numbering of this response (Tables R1–R7, with R2 and R5 in the main text and the others in Appendix C). Items that need new experiments or renderings (three seeds, the cumulative ablation from LiNeXt, the new insets in Fig. 5, and cross-sensor evaluation) will be added in the final version. Tables are numbered consistently across our responses to all reviewers.

W1/Q1 — Why the full FASSE and FRVO configurations

We agree that Table 2(c,d) shows trade-offs rather than uniform gains. Like the hyperparameters (W2), the full configuration follows from the footprint model and was fixed before any evaluation; the internal ablations were run afterwards to measure each choice, not to select one. CD and JSD BEV grow with how far points lie from the reference surface. Voxel IoU at 0.1 m scores each voxel only as hit or miss, regardless of how far a point outside the reference voxels lies from the surface. The two groups of metrics therefore weight coverage and placement precision differently. Against LiNeXt, BeamScene improves both groups, i.e., all six metrics. Neither uniform scale weighting (“w/o footprint routing” in Table 2c) nor CD + FT improves all six metrics over the full model, or vice versa; the full configuration lies at the CD and JSD BEV end of this trade-off.

The simpler variants are settings of the same design. Uniform scale weighting is the footprint gate in the limit τs→∞\tau_s \to \infty. CD + FT is the full objective with λray=0\lambda_{\text{ray}} = 0. Kernel sharing reduces the parameters of the complete model by 41.80% relative to the same backbone with independent kernels (Appendix A). Among these settings, the full configuration is the one with the lowest CD and JSD BEV.

Why uniform scale weighting gains fine IoU. Under the gate, a return at 10 m (si=0.05s_i = 0.05 m) puts less than 0.1% of its weight on the 0.64 m and 1.28 m scales. Uniform scale weighting gives these two scales 25%. We attribute the higher fine IoU to this coarse context, which fills space more aggressively but places points less precisely. This is consistent with its worse CD (0.207 vs. 0.205) and JSD BEV (0.335 vs. 0.332).

Why the ray term is kept. The ray term (first-return term plus one-sided free-space penalty) lowers CD from 0.210 to 0.205 and JSD BEV from 0.336 to 0.332. The first-return term alone does not (0.211 CD), so the gain requires the free-space penalty. These are the two metrics most sensitive to points in front of observed surfaces, which act as phantom obstacles in driving.

Where the full configuration is preferable. The gate and uniform scale weighting differ mainly for near and mid-range returns. At 80 m the gate already puts 43% of its weight on the 0.64 m and 1.28 m scales. The full configuration is therefore preferable where surfaces are densely sampled and precise placement matters, such as nearby road boundaries and facades. Uniform scale weighting, which adds coarse context everywhere, favours filling larger gaps.

The comparison with LiNeXt does not depend on this choice. Every variant in Table 2, including those with a whole component removed, exceeds LiNeXt at 0.2 m and 0.1 m IoU (at least 23.61 and 7.68, vs. 19.45 and 6.30). This also means that no single component accounts for the whole gain over LiNeXt. We do not attribute it entirely to the three components, and the final version adds a cumulative ablation that starts from LiNeXt retrained in our implementation.

Revision. In the revised PDF, Sec. 4.3 states how the configuration was selected and presents uniform scale weighting and CD + FT as alternative operating points, and Sec. 4.1 no longer calls fine IoU the most direct measure of recovering unobserved regions. The fuller explanation above and the statement in Sec. 3.3 that the effective receptive field depends on the footprint will be added in the final version.

W2/Q2 — Hyperparameters, tuning split, common configuration, sensitivity

No hyperparameter was tuned on sequence 08. Each was fixed before any evaluation, either taken from prior work or set from its role in the equations of Sec. 3. Table 4 of the revised PDF lists all of them, including five values the submission omitted (smin⁡s_{\min}, εot\varepsilon_{\text{ot}}, ℓmin⁡\ell_{\min}, λf\lambda_f, β\beta).

AuthorsOpenReview ↗

Response to Reviewer kQbf (2/5)

W2/Q2 — Hyperparameters, tuning split, common configuration, sensitivity (cont.)

Taken from prior work. The data sizes N=18,000N = 18{,}000, M=180,000M = 180{,}000, and U=6U = 6 follow LiDiff and LiNeXt (Appendix A). BFCC replaces LiNeXt's manually defined range groups and fixed-variance Gaussian noise (Sec. 1) with a deterministic rule, so those choices no longer exist.

Candidate construction (BFCC, Sec. 3.2).

  • Δθ,Δϕ=0.20∘,0.42∘\boldsymbol{\Delta\theta, \Delta\phi = 0.20^\circ, 0.42^\circ}, the nominal sensor resolution (Appendix A). The allocation does not depend on them: Eq. (1) gives Ai=ri2 Δθ ΔϕA_i = r_i^2\,\Delta\theta\,\Delta\phi, and both the quantile clip and the normalisation wi=Aˉi/∑jAˉjw_i = \bar A_i / \sum_j \bar A_j cancel the constant Δθ Δϕ\Delta\theta\,\Delta\phi. Apart from the floor smin⁡s_{\min}, an error in these values only rescales sis_i. A 20% error in Δθ Δϕ\Delta\theta\,\Delta\phi rescales sis_i by 10% and shifts log⁡si\log s_i by 0.09, about one eighth of the spacing between gate scales.
  • γ=1\boldsymbol{\gamma = 1} makes the number of candidates proportional to the footprint area, which is the design goal of BFCC.
  • qℓ,qu=0.05,0.95\boldsymbol{q_\ell, q_u = 0.05, 0.95}. AiA_i grows monotonically with range, so the clip affects only the nearest and the farthest 5% of returns.
  • nmax⁡=24\boldsymbol{n_{\max} = 24}. Prop. 1 requires N≤M≤N nmax⁡N \le M \le N\,n_{\max}, i.e., nmax⁡≥M/N=10n_{\max} \ge M/N = 10. A return receives about 1+(M−N) wi=1+9 Aˉi/mean⁡(Aˉ)1 + (M - N)\,w_i = 1 + 9\,\bar A_i / \operatorname{mean}(\bar A) candidates, so nmax⁡n_{\max} binds only for returns whose clipped area exceeds 23/9≈2.5623/9 \approx 2.56 times the mean.
  • κ=0.35\boldsymbol{\kappa = 0.35}. The outermost candidate lies at radius κ si\kappa\, s_i (Eq. 2). Since Ai=si2A_i = s_i^2, this radius is below that of the disc of equal area, si/π≈0.56 sis_i/\sqrt{\pi} \approx 0.56\, s_i, so all candidates stay inside the footprint. With the values above it also equals half the horizontal spacing: 0.35Δθ Δϕ=0.101∘≈Δθ/20.35\sqrt{\Delta\theta\,\Delta\phi} = 0.101^\circ \approx \Delta\theta/2.
  • smin⁡=0.01\boldsymbol{s_{\min} = 0.01} m equals the finest scale g1g_1. By Prop. 3, a return with sis_i below g1g_1 already gives its largest weight to g1g_1, so the floor only removes the divergence of log⁡si\log s_i as r→0r \to 0 and leaves the routing unchanged. It is active only below r≈2r \approx 2 m, where the expansion radius is κ smin⁡=3.5\kappa\, s_{\min} = 3.5 mm. The allocation does not use smin⁡s_{\min}; the quantile clip bounds AiA_i separately.
  • ξ=0.754877666\boldsymbol{\xi = 0.754877666}, the inverse of the plastic number. It is the irrational offset of Eq. (2) that decorrelates the starting azimuth of neighbouring returns; the golden angle π(3−5)\pi(3 - \sqrt{5}) is the step between candidates.

Scale routing (FASSE, Sec. 3.3).

  • gk=0.01⋅2k−1\boldsymbol{g_k = 0.01 \cdot 2^{k-1}} m, K=8\boldsymbol{K = 8}. The grid is dyadic; the finest scale, 0.01 m, equals sis_i at r≈2r \approx 2 m, and the coarser scales provide context at long range.
  • τs=0.70≈ln⁡2=0.693\boldsymbol{\tau_s = 0.70 \approx \ln 2 = 0.693}, the spacing of log⁡gk\log g_k. By Eq. (3) and Prop. 3, a return at a grid scale gives each adjacent scale 61% of the peak weight and scales two steps away 14%, so it draws mainly on its own and the adjacent scales.
AuthorsOpenReview ↗

Response to Reviewer kQbf (3/5)

W2/Q2 — Hyperparameters, tuning split, common configuration, sensitivity (cont.)

Visibility objective (FRVO, Sec. 3.5).

  • δf=τr=0.10\boldsymbol{\delta_f = \tau_r = 0.10} m, the size of the finest evaluation voxel (0.1 m, Sec. 4.1).
  • Frustum cell 5∘×3∘×5\boldsymbol{5^\circ \times 3^\circ \times 5} m and ray bin 1∘×1∘\boldsymbol{1^\circ \times 1^\circ}. In units of the sensor resolution these are 25×725 \times 7 and 5×2.45 \times 2.4 angular steps, so each cell and each ray spans several beams.
  • λrad=0.25\boldsymbol{\lambda_{\text{rad}} = 0.25}. In Eq. (5), for a depth offset dd the linear term 0.25 d/ℓc0.25\, d/\ell_c exceeds the quadratic term (d/ℓc)2(d/\ell_c)^2 exactly when d<0.25 ℓcd < 0.25\,\ell_c. The radial term therefore acts where the quadratic cost is flat, which is how it penalises depth disagreement at small offsets.
  • εot=0.02\boldsymbol{\varepsilon_{\text{ot}} = 0.02}, ℓmin⁡=0.1\boldsymbol{\ell_{\min} = 0.1} m. Eq. (5) normalises costs by ℓc\ell_c, so εot\varepsilon_{\text{ot}} is dimensionless and applies unchanged at every range. Its entropic blur, ℓcεot≈ℓc/7\ell_c\sqrt{\varepsilon_{\text{ot}}} \approx \ell_c/7, matches the spacing of the at most 64 points per cell (Appendix A), about ℓc/8\ell_c/8. ℓmin⁡\ell_{\min} equals the finest evaluation voxel, like δf\delta_f and τr\tau_r; since ℓc=0.068 rˉ\ell_c = 0.068\,\bar r, it is active only for cells with rˉ<1.5\bar r < 1.5 m.
  • λf=0.5\boldsymbol{\lambda_f = 0.5} gives the free-space penalty of Eq. (6) the same form as the quadratic zone of SmoothL1, 0.5 x20.5\,x^2. A prediction more than δf\delta_f in front of the reference first return is therefore pulled back about twice as strongly as one behind it. Table 2(d) reports λf=0\lambda_f = 0 (CD + FT + first-return).
  • λFT=λray=0.02\boldsymbol{\lambda_{\text{FT}} = \lambda_{\text{ray}} = 0.02} keep FRVO auxiliary to the Chamfer loss; Table 2(d) reports each at zero. The six Sinkhorn iterations are a computational setting with bounded memory (Appendix A).

Decoder (Sec. 3.4). β=0.1\boldsymbol{\beta = 0.1} m. In Eq. (4), rm,u−cm=βtanh⁡ ⁣(Dr(⋅))r_{m,u} - c_m = \beta \tanh\!\big(D_r(\cdot)\big) with tanh⁡\tanh applied element-wise, so every coordinate of a child offset lies in (−β,β)(-\beta, \beta) and its length is below 3 β≈0.17\sqrt{3}\,\beta \approx 0.17 m. Since β\beta equals the finest evaluation voxel, a child can cross at most one 0.1 m voxel boundary per axis, i.e., reach the voxels adjacent to its parent's but no farther. The refinement therefore stays local.

Tuning split. As in LiDiff and LiNeXt, we train on sequences 00–07 and 09–10 and report results on sequence 08, since the test labels are not public. Sequence 08 was never used to select a hyperparameter or a checkpoint: we always report the final epoch.

Common configuration. SemanticKITTI and KITTI-360 use the same configuration and the same checkpoint, without fine-tuning.

Sensitivity. The internal ablations in Table 2 already set individual hyperparameters to their extremes. Table R6 reads them as a sensitivity analysis, as changes relative to the full model (CD 0.205, JSD BEV 0.332, [email protected] 8.74).

Table R6. Effect of each hyperparameter at its extreme setting (values from Table 2).

Setting in Table 2HyperparameterΔ\DeltaCDΔ\DeltaJSD BEVΔ\Delta[email protected]
CD + RayλFT=0\lambda_{\text{FT}} = 0+0.011+0.011+0.024+0.024−0.39-0.39
Isotropic 3D expansiontangent plane+0.007+0.007+0.011+0.011−0.32-0.32
Distance-grouped allocationγ,qℓ,qu\gamma, q_\ell, q_u+0.006+0.006+0.009+0.009−0.06-0.06
CD + FT + first-returnλf=0\lambda_f = 0+0.006+0.006+0.008+0.008−0.09-0.09
CD + FTλray=0\lambda_{\text{ray}} = 0+0.005+0.005+0.004+0.004+0.78+0.78
Uniform scale weightingτs→∞\tau_s \to \infty+0.002+0.002+0.003+0.003+0.70+0.70

λFT\lambda_{\text{FT}} has the greatest influence on CD and JSD BEV. τs\tau_s and λray\lambda_{\text{ray}} mainly trade fine IoU against CD and JSD BEV, as discussed under W1. The allocation rule mainly affects the BEV layout. Even at these extremes, CD changes by at most 0.011, and every setting still exceeds LiNeXt at [email protected] and [email protected] (at least 24.31 and 8.35, vs. 19.45 and 6.30), although CD + Ray falls behind LiNeXt in CD (0.216 vs. 0.214).

Reviewer_9H4nOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a forward approach for Lidar scene completion, coined BeamScene. Instead of arbitrary hyperparameter choices, it considers lidar sensors spefications to make sensor-informed choices, and to construct a sensor-informed pipeline. Specifically, similar to the recent LiNeXt, the authors revisit the choice of architecture and suggest a feedforward approach instead of the more recent diffusion-based architectures. The proposed method, then construct a dense candidate point cloud by deduplicating the sparse input lidar following the expected beam footprint in the physical world, and use lighter multiscale sparse encoder with coarse and fine decoders. The method is evaluated on SemanticKITTI and KITTI-360, showing slightly better performance that the best competitor. A thorough ablation shows that each component contribute (sometimes very midly) to the overall improvement.

Critical strengths

The paper topic is of interest, and the proposed sensor-informed formulation is interesting. While sensor-informed completion is already well explored, these methods often struggle to compete with the best ones, and I appreciate the performance of the proposed method.

While experiments scope is narrow, the ablations are thorough which demonstrate the validity of all the proposed components though they sometimes bring marginal improvement (arguably within noise ?). The method seems to bring some quantitative improvement on the popular SemanticKITTI dataset.

Critical weaknesses

The paper writing is not smooth, and required some efforts for understanding such that it may hinder the understanding of the proposed method and maybe its evaluation. Some of the statements are also overclaimed given the current state of the literrature and the contribution lines with respect to the recent LiNeXt are blurred. Below are the main criticisms.

W1. Contribution wrt LiNeXt is unclear:

  1. Presentation: Figs. 1 and 2 compare BeamScene to diffusion-based methods, claiming much faster inference due to the single forward pass. However, LiNeXt already proposed forward pass completion, and the boost seems negligible with respect to LiNeXt.
  2. Methodological contribution: It isn't clear which part of the method is directly taken from LiNeXt and which is new. For example: the FASS module (Sec. 3.2) seems similar to the MSCC module of LiNeXt, also relying on multi-scale sparse convolution. Similarly, the decoders (Sec. 3.4) seems to resemble the Noise to Coarse Module and Refinement Module of LiNeXt. I would appreciate clarification of the contributions (cf. Authors questions).
  3. Performance contribution: In several places the gap with diffusion-based methods is largely highlighted (Figs. 1, 2, 4, etc.) but if the method builds on LiNeXt -- even conceptually -- the gap with LiNeXt should instead be highlighted as this is the true experimental contribution of the paper. From Fig. 4a, 4b, 4c, and Tab. 1 the refined improvement is very minor and not really noticeable in the visuals.

W2. The BFCC (Sec. 3.2) spreads the candidate points a beam-orthogonal plane which is a valid hypothesis. However, it should be highlighted that the surface orientation may not be orthogonal, this is largely the case for the ground for example. It would be important to clarify/quantify the implication of such approximation.

W3. Some statements are somehow overclaimed:

  1. line 192: "where the measurement is already reliable" => Lidar sensors have also a depth measurement error (also available in the specs sheet) which could be exploited here.
  2. Sec. 3.5 introduces a frustum-ray objective building on the assumption that no points can lie between a return and the sensor. While this is true on a single scan basis, ground truth in SemanticKITTI/KITTI-360 and other autonomous driving datasets are built by aggregating multiple scans. In that context, such objective may be incorrect. Besides, the formulation assumes the scene static and becomes invalid considering moving objects.

W4. The current method positions in the line of the "sensor-informed" approaches. A number of relevant sensor-informed works already exist for completion/reconstruction objects, but are omitted in the literature review: (OctoMap, Autonomous Robots 2013; Roldão et al., arXiv 2018; Sulzer et al., ICPR 2022; etc.).

W5. Qualitative results: In Fig. 5, the visual comparison to LiNeXt is not trivial. Besides the caption highlighted differences in Fig. 6 appear rather speculative.

W6. Why showing LiDiff in Figs. 1,5,7,8,9 instead of LiDPM since it shows better performance ?

W7. The paper writing is technical and lack research intuitions. In particular, the introduction should broadly present the method before technical details. The modules names BFCC/FASSE/FRVO are also not intuitive and hard to associate to their functional behavior. Fig. 3 could include the module names for easier grasping at first sight.
W8. On experimental evaluation: KITTI-360 is reported for generalization. However, KITTI-360 is a dataset very similar to SemanticKITTI: recorded in the same districts, with similar (but not identical) sensor setup, and around the same years (despite the delayed publication of KITTI-360). Hence, I have doubt about the ability of KITTI-360 to reflect generalization. Instead, nuScene or Waymo are often used to demonstrate generalization and could have been included in the paper for stronger generalization assessment.

W9. Formatting: It is not clear why BFCC ablation is additive (eg, "w/ isotropic 3D expansion") instead of substractive (eg, FASSE "w/o kernel sharing") like the rest of the table.

Hornung, A., Wurm, K. M., Bennewitz, M., Stachniss, C., & Burgard, W. (2013). OctoMap: An efficient probabilistic 3D mapping framework based on octrees. Autonomous robots, 34(3), 189-206.
Roldão, L., De Charette, R., & Verroust-Blondet, A. (2018). A statistical update of grid representations from range sensors. arXiv preprint arXiv:1807.08483.
Sulzer, R., Landrieu, L., Boulch, A., Marlet, R., & Vallet, B. (2022, August). Deep surface reconstruction from point clouds with visibility information. In 2022 26th International Conference on Pattern Recognition (ICPR) (pp. 2415-2422). IEEE.

Questions for authors

Please refer to the weaknesses for remaining concerns. In particular I would appreciate answers on (but not restricted to) the following:

  • How does the FASSE (Sec. 3.3) differ from the MSCC module of LiNeXt ? In particular, both rely on multi-scale sparse convolution, so what is the difference in the problem setup and the technical implementation ?
  • How the decoders (Sec. 3.4) differ from the noise-to-coarse and refinements modules of LiNeXt ? (cf. Weaknesses)
  • How the method positions wrt the other sensor-informed works in reconstruction / completion ? (cf. Weaknesses)
7 comments
AuthorsOpenReview ↗

Response to Reviewer 9H4n (1/7)

We thank the reviewer for the careful review and for recognising the performance of the method and the thoroughness of the ablations. The central concern is the delineation of our contribution with respect to LiNeXt. We agree that the submission does not make it explicit enough, and this response states precisely what is inherited and what is new. Several of the analyses below were prepared during the discussion and are shared across reviewers; tables are numbered consistently across all our responses. We have uploaded a revised PDF in which all changes are marked in blue; the new tables keep the numbering of this response (Tables R1–R7, with R2 and R5 in the main text and the others in Appendix C). Items that need new experiments or renderings (three seeds, the cumulative ablation from LiNeXt, the new insets in Fig. 5, and cross-sensor evaluation) will be added in the final version.

Summary. (1) BeamScene is built on LiNeXt; Table R1 lists exactly what is inherited (single-pass pipeline, voxel pyramid, N2C/Refine decoders) and what is new (footprint-based candidate construction with exact anchors, a shared-kernel encoder with a parameter-free footprint gate, and a training-only visibility objective). (2) We no longer present runtime as a contribution over LiNeXt: the main result is the coarse-stage gain (+5.46+5.46 / +2.44+2.44 IoU at 0.2 / 0.1 m), concentrated in unobserved and long-range regions (Table R2), and the refined stage is stated as parity or better with 41.4% fewer parameters. (3) We correct the overstated wording (W3), add the missing sensor-informed references (W4), and describe KITTI-360 as cross-sequence rather than cross-sensor evaluation (W8). (4) In the revised PDF, LiDPM replaces LiDiff in Figs. 1, 5, 7 and 8 (W6), and Sec. 2 now cites Hornung et al. (2013, OctoMap), Roldão et al. (2018) and Sulzer et al. (2022) (W4).

W1 / Q1 / Q2 — Contribution with respect to LiNeXt

What is inherited and what is new. BeamScene is built on the LiNeXt pipeline. The revised PDF states this in Appendix C, which contains Table R1 below.

Table R1. Components inherited from LiNeXt and components introduced by BeamScene.

StageInherited from LiNeXtNew in BeamScene
Candidatessingle-pass expansion of the scan to M=180M=180k pointsBFCC replaces replication by manually defined range groups (range quartiles, 5/8/12/15 copies) and fixed-variance Gaussian noise on every copy: bounded r2r^2-proportional budget; deterministic tangent-plane expansion (golden angle), so inference is deterministic, whereas LiNeXt samples new noise for every scan; immutable anchors; every candidate carries its parent's footprint sis_i
Encoderdyadic voxel pyramid 0.01⋅2k−10.01\cdot2^{k-1} m; two residual submanifold convolutions per resolution; applied to scan and candidates. In MSSC these weights are separate per resolution and per branch (two MSSC instances), and the resolutions are fused by concatenation + MLPone kernel pair shared across all eight resolutions and both branches (−41.8%-41.8\% parameters, App. A); per-point, parameter-free footprint gate instead of concatenation + MLP
Seeds / coarse decoder (N2C)serialised local aggregation, progressive downsampling, seed-conditioned decoding; a residual is predicted and subtracted for every pointanchor mask on the residual (Eq. 4): measured returns are kept bit-exactly (Prop. 2)
Refinement (Refine)SPD-style upsampling, U=6U=6, tanh⁡\tanh-bounded child offsets, stage-wise training with the coarse network frozenbound β=0.1\beta=0.1 m (1 m in LiNeXt); first child of each anchor kept exactly
SupervisionChamfer distanceFRVO (frustum-local transport, first-return term, free-space penalty), training only

The new components are linked by a single quantity, the footprint sis_i, the equivalent length scale of the area AiA_i that one beam covers at range rir_i (Eq. 1). It sets the candidate budget and the expansion radius in BFCC, is inherited by every candidate, and selects the scale in the encoder gate. FRVO adds the visibility counterpart of the same acquisition model. This shared derivation, rather than any single module, is what we mean by a physically consistent pipeline.

Single-pass inference is LiNeXt's contribution, not ours. We agree with the reviewer. We will revise Figs. 1–2 and the introduction so that LiNeXt, not diffusion-based completion, is the primary reference, keeping the diffusion results as context. We will no longer present runtime as a contribution relative to LiNeXt: the two are on par (0.234 s vs. 0.245 s; 0.41 s vs. 0.43 s with refinement).

AuthorsOpenReview ↗

Response to Reviewer 9H4n (2/7)

W1 / Q1 / Q2 — Contribution with respect to LiNeXt (cont.)

The ablations already isolate each departure from LiNeXt under identical training. Several rows of Table 2 replace one of our components with the corresponding LiNeXt design, or a simpler variant of it, while keeping everything else, including the training schedule, fixed:

  • Candidate construction. LiNeXt's construction has two parts, and each is isolated separately. “Distance-grouped allocation” replaces footprint allocation with LiNeXt's range groups (range quartiles with 5/8/12/15 copies). “Isotropic 3D expansion” replaces the tangent-plane expansion with isotropic 3D expansion (Table 2b). “w/o BFCC” replaces both, with uniform replication and fixed-variance perturbation (Table 2a). Since the anchors are produced by BFCC (Alg. 1), this row also has no anchors.
  • Encoder. “w/o FASSE” uses independent per-resolution kernels with concatenation, i.e., the MSSC design (Table 2a).
  • Supervision. “w/o FRVO” trains with the Chamfer loss only, as in LiNeXt (Table 2a).

Each of these substitutions worsens all six metrics. Appendix C of the revised PDF states this mapping explicitly, and Table 2(b) now uses the subtractive labels (W9).

Where the gain over LiNeXt comes from. Table R2 splits voxel IoU at 0.2 m (the metric with the largest gain) by visibility from the input viewpoint and by range, using the public LiNeXt checkpoint and our trained checkpoints. A reference point is observed if it lies within 0.3 m of the input scan's first return along its ray, and unobserved if it lies behind that return or on a ray with no return. Voxels are assigned to a region by their centre.

Table R2. Voxel IoU at 0.2 m on SemanticKITTI by visibility and range (no refinement).

ObservedUnobserved0–20 m≥\ge 20 m
Share of reference points (%)13.186.931.069.0
LiNeXt63.9015.8731.8114.53
BeamScene67.4021.5935.0219.78

The gain is +5.72+5.72 points (+36% relative) on unobserved voxels and +5.25+5.25 (+36%) beyond 20 m, against +3.50+3.50 on observed voxels. It is therefore concentrated where the scan provides the least evidence.

What the single-component ablations explain, and what they do not. Each row of Table 2(a) removes one component and keeps the other two. It measures the marginal contribution of that component given the others. It is not a path from LiNeXt, and the marginal contributions need not add up to the gain over LiNeXt. Table R3 compares the two.

Table R3. Gain of BeamScene over LiNeXt, and marginal contribution of each component (full model minus its single-component removal, Table 2a). Coarse stage, SemanticKITTI.

MetricGain over LiNeXtBFCCFASSEFRVOSumShare of the gain
[email protected]+5.46+5.46+1.02+1.02+0.59+0.59+1.30+1.30+2.91+2.9153%
[email protected]+2.44+2.44+1.06+1.06+0.47+0.47+0.75+0.75+2.28+2.2893%
  • At 0.1 m, the marginal contributions add up to nearly the whole gain (2.28 of 2.44 points), and BFCC contributes the most.
  • In CD, removing BFCC or FRVO alone loses the whole advantage over LiNeXt (0.216 and 0.224, vs. 0.214 for LiNeXt and 0.205 for the full model). The CD gain therefore requires both.
  • At 0.2 m, the marginal contributions account for 2.91 of the 5.46 points. Single removals cannot attribute the remaining 2.55, and Table 2 cannot separate two explanations. First, the components may overlap at this resolution: when one is removed, the other two recover part of its effect, so each marginal contribution understates what the component adds on its own. Second, the remainder may come from what all three rows share and LiNeXt does not. The anchors are not among these: they are produced by BFCC, so “w/o BFCC” has none, and their effect is part of the BFCC contribution. What all three rows share is our implementation and training of the inherited backbone, whereas the LiNeXt row is its public checkpoint.

We therefore do not claim that the three components explain the whole gain at 0.2 m. We keep the claims that Table 2 supports: each component improves all six metrics given the others, the marginal contributions add up to nearly the whole gain at 0.1 m, and the CD gain requires both BFCC and FRVO. For the final version, we will add a cumulative ablation that starts from LiNeXt retrained in our implementation and adds BFCC, FASSE and FRVO in turn, with the anchors as a separate step within BFCC. It separates the two explanations and replaces the marginal view with a path from LiNeXt.

AuthorsOpenReview ↗

Response to Reviewer 9H4n (3/7)

W1 / Q1 / Q2 — Contribution with respect to LiNeXt (cont.)

The refined-stage gap is small, and we will say so. We agree that with refinement the improvement over LiNeXt† is minor (CD 0.149 →\to 0.144, [email protected] 15.09 →\to 15.85). The inherited 6×6\times refinement largely closes the coverage gap of the coarse stage. A paired block bootstrap over sequence 08 (10,000 resamples, blocks of 100 frames) gives intervals that exclude zero: CD −0.005-0.005 [−0.0083-0.0083, −0.0022-0.0022] and [email protected] +0.68+0.68 [+0.55+0.55, +0.81+0.81]. We have nevertheless revised the claim at this stage to parity or better with 41.4% fewer parameters (1.23M vs. 2.10M), rather than “best on all six metrics” alone. In the revised PDF, the abstract, the contribution list and Sec. 4.2 place the main result at the coarse stage, where the gains are +5.46+5.46 / +2.44+2.44 points at 0.2 / 0.1 m.

Q1 — FASSE vs. MSSC. The problem setup and the voxel pyramid are the same as in MSSC, as Table R1 (Appendix C) now states. There are two technical differences.

  • Kernel sharing. MSSC uses separate sparse kernels and a separate embedding MLP for each resolution (spconvk,1\mathrm{spconv}_{k,1}, spconvk,2\mathrm{spconv}_{k,2}, MLPk\mathrm{MLP}_k). In the released LiNeXt code, the scan and the noisy candidates are also encoded by two separate MSSC instances. FASSE ties one pair of kernels across all eight resolutions and both input branches. Features from different resolutions are then produced by the same local operator and are directly comparable, and the complete model uses 41.8% fewer parameters than the same backbone with independent kernels.
  • Fusion. MSSC concatenates the resolutions and applies an MLP, i.e., the same learned fusion for every point. FASSE takes a per-point weighted sum whose weights depend only on the footprint sis_i (the footprint gate), which every candidate inherits from BFCC. The preferred scale therefore follows the sampling density of each point.

“w/o FASSE” in Table 2(a) is the MSSC design within our pipeline, and Table 2(c) separates the two differences. Weight sharing across receptive fields has precedent, e.g., TridentNet (Li et al., ICCV 2019), which the revised PDF now cites in Sec. 3.3. Our contribution is the parameter-free, footprint-driven scale selection in sparse scene completion, i.e., range-adaptive routing whose offset comes from the sensor model instead of tuning. Compared with uniform scale weighting (“w/o footprint routing” in Table 2c), the footprint gate trades some fine-scale IoU for lower CD and JSD BEV (Table 2c).

Q2 — Decoders vs. N2C and Refine. The seed encoder, the seed-conditioned coarse decoder and the SPD-style refinement follow LiNeXt, as Table R1 now states. This includes the residual form of the coarse output and the tanh⁡\tanh bound on the refinement offsets. What differs is how measured returns are treated.

  • In LiNeXt, every copy of a return, including the copies that stand for the measurement itself, receives Gaussian noise. A residual is then predicted and subtracted for every point, and every child is displaced. Measured returns are therefore not guaranteed to appear in the output.
  • In BeamScene, the anchor mask of Eq. (4) selects the stored coordinate instead of the residual, and the first child of each anchor is kept exactly. Every measured return therefore appears bit-exactly in both outputs (Prop. 2). We also set the offset bound to β=0.1\beta=0.1 m instead of 1 m, so that refinement stays local.

We do not claim the anchors themselves as a source of the metric gains. The 18,000 anchors are 10% of the coarse output and under 2% of the refined output, and Table R2 shows that the gain lies mainly in unobserved regions. The anchor mask is first a structural guarantee: by Prop. 2, the missing-first-return rate of BeamScene on the input rays is exactly zero. Whether keeping the measurements fixed also helps the decoder place the other points is not isolated by Table 2: the anchors are produced by BFCC, so their effect is contained in the BFCC contribution but not separated from it (W1).

Comments on the paper

1 comment

Dear AC and reviewers,

We sincerely thank the AC and the three reviewers for their time and careful engagement throughout the review and discussion. We took every question seriously and answered each point in detail.

Method and novelty. We propose BeamScene, a new physically consistent framework for single-pass LiDAR scene completion. It addresses a key problem of existing single-pass methods: candidate placement, feature scale, observation retention, and visibility supervision are each governed by unrelated hand-set rules. BeamScene derives all four from one physical quantity, the beam footprint of each return, through (i) BFCC, a footprint-proportional candidate constructor with tangent-plane expansion and bit-exact retention of every measured return; (ii) FASSE, a shared-kernel multi-scale encoder with a parameter-free footprint gate (41.8% fewer parameters); and (iii) FRVO, a training-only first-return and free-space objective at no inference cost.

Results. On SemanticKITTI, BeamScene improves on the strongest single-pass baseline, LiNeXt, in all six metrics with 43.7% fewer parameters, raising voxel IoU by +5.46 / +2.44 at 0.2 / 0.1 m, mainly in unobserved and distant regions (Table R2). With refinement, it matches or exceeds all compared methods on all six metrics of SemanticKITTI and KITTI-360 (without fine-tuning) with 41.4% fewer parameters, and it is 17–99× faster than diffusion-based completion.

Additional information provided in response to the reviewers. To address the reviewers' questions, we added: an explicit comparison with LiNeXt (Table R1) and an attribution of the gain (Table R3); a visibility and range breakdown (Table R2); a direct free-space evaluation on unseen raw-scan rays (Table R5); a bound for the tangent-plane approximation (Table R4) and for the dilution case raised by a reviewer; the full hyperparameter list, a sensitivity analysis (Table R6), and paired confidence intervals (Table R7); the missing sensor-informed related work (OctoMap, Roldão et al., Sulzer et al., and others); corrected wording on our claims; and LiDPM in Figs. 1, 5, 7 and 8. All changes are marked in blue in the revised PDF. We believe these additions address all concerns raised by the reviewers.

We thank the reviewers again for their insightful questions, which have made the paper clearer and stronger.

Best regards,

All authors of the paper

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.