acceptodds
Withdrawn from ICLR 2027

ADC: Lidar Localization for Autonomous Driving in Various Weather Conditions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate and robust 6-DoF localization is essential for autonomous vehicles, where LiDAR has emerged as a mainstream alternative to vision-based sensors due to its direct 3D geometric perception. However, existing LiDAR localization methods are primarily designed for clear weather, overlooking severe signal degradation in adverse environments. In rain and snow, precipitation-induced backscattering, laser attenuation, and snow accumulation lead to airborne clutter, point dropouts, and structural corruption in LiDAR observations, severely degrading localization performance. To break this bottleneck, this paper presents ADC, the first map-free localization framework specifically designed for robust LiDAR localization under adverse weather conditions. To combat weather-induced noise and observation inconsistency, we propose a hierarchical sparse encoding architecture coupled with perspective-invariant global aggregation and an adaptive point-wise feature recalibration module to effectively filter out airborne clutter while amplifying stable rigid structures. Furthermore, we introduce a local geometric rigidity constraint that enforces an implicit structural skeleton on degraded point clouds, resolving coordinate drift and preserving topological consistency. Extensive evaluations on the challenging Boreas benchmark and the standard Oxford RobotCar dataset demonstrate that ADC achieves state-of-the-art performance. Specifically, our method reduces mean translation error by 70% on the adverse-weather Boreas dataset and by 18% on the Oxford benchmark relative to the strongest competing map-free baselines, demonstrating robust and weather-resilient localization across diverse driving conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Localization, 3D perception Localization, Adverse-weather

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gtUMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets map-free LiDAR relocalization in rain and snow. A sparse voxel encoder runs on four rotations of each scan, a 25-way regional prior from clustered training positions is concatenated to the point features, a per-point channel gate rescales them, and a pairwise distance term regularises the predicted scene coordinates. Training uses four clear Boreas traversals. Testing uses five traversals, three snow and one rain, plus Oxford RobotCar. Reported mean errors are 0.98 m and 0.89° on Boreas, 0.59 m and 0.90° on Oxford.

Critical strengths
  1. The design maps one mechanism onto each failure mode named in the introduction. The sharpest is Eq. 13, which constrains pairwise distances while L_coord treats each point independently.

  2. The setting is new: map-free regression trained on clear traversals only, then tested on snow and rain of the same route. Sub-metre on all five sequences, 70% below the strongest baseline, with no map.

Critical weaknesses
  1. No module ablation separates snow from sun. Table 5 gives one Boreas column, the mean over all five test traversals, so it cannot show that APFR or EdgeLoss helps under weather. The lead over TempLoc is 69.3% on the sunny traversal, 71.2% on the snow traversals and 68.1% on the rainy one. The rainy traversal is also easier than all three snow ones for every method in Table 1. Appendix A concedes the split is not an isolated measurement of precipitation.

  2. The baselines were re-trained, and the protocol is missing. TempLoc's own paper evaluates on NCLT and Oxford RobotCar, not Boreas, so these are re-implementations. RALoc reads 1.24° on Oxford and 3.74° on Boreas, and GTRLoc 1.03° against 7.40°, both worse than PosePN++ at 1.07°. No validation split or variance is reported either, which leaves the 0.13 m Oxford margin and the 0.02 m APFR gain inside run-to-run noise.

  3. The degradation the method targets is never measured. No point counts, clutter fraction, per-point coordinate error or RANSAC inlier ratio appears. Figure 9's caption calls its segment an illustrative case. Table 6 compounds this. The single-viewpoint row reads 5.09 m against 2.50 m for the vanilla baseline. EdgeLoss also normalises by the number of valid pairs, so it weakens on exactly the scans with the fewest.

Questions for authors
  1. Could you report the ablation per weather condition? Table 5 gives one mean over all five traversals.

  2. Does the single-viewpoint row of Table 6 include APFR and EdgeLoss? At 5.09 m it is worse than the vanilla baseline at 2.50 m, yet the last row equals the full model, so the row setup is unclear.

  3. Section 4.1 implies the baselines were re-trained, but no configuration appears. Please give each one's epochs, learning-rate search and voxel settings, plus the Oxford training sequences and its error on a clear-weather Boreas split.

  4. RALoc reads 1.24° on Oxford and 3.74° on Boreas, GTRLoc 1.03° against 7.40°. Rotation is the easy part, since HypLiLoc reaches 0.94° with 4.31 m translation, which suggests incomplete training. Were both adapted to the Boreas protocol?

  5. No validation split is described. On which data were λ_e, τ, the rotation view count and the 25 regional clusters selected?

  6. The Oxford margin over TempLoc is 0.13 m. Please report mean and standard deviation over three seeds for Table 2.

  7. Section 2.3 is the only subsection on adverse-weather localization and cites nothing. Which prior adverse-weather localization work does the "first" claim in the abstract position itself against, and is any of it map-free?

  8. Table 1 reports LighterBEV but Table 4 reports Bevplace++ as the second map-based method, so the two tables cover different systems and I cannot tell whether ADC also outranks Bevplace++ in accuracy. Could both appear in both?

Reviewer_GD77OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ADC is a scene coordinate regression framework for LiDAR relocalization across weather conditions. It combines sparse local features, global pooling over four canonical yaw rotations, a regional response supervised by clusters of training positions, point-wise channel gating (APFR), and a local pairwise-distance consistency loss (EdgeLoss). RANSAC estimates the final 6-DoF pose.

The paper reports mean translation errors of 0.98 m on Boreas and 0.59 m on Oxford, improving over TempLoc by approximately 70.2% and 18.1%. Frozen-model synthetic-corruption experiments also show lower absolute errors and meaningful improvements in median-error robustness.

Critical strengths
  1. Substantial reported accuracy gains across two datasets. ADC obtains 0.95–1.00 m mean translation error across the five Boreas test traversals and improves over the listed SCR baselines on Oxford. The snowy 2020-12-01 traversal is particularly encouraging: LightLoc increases from 2.78 m on the sunny traversal to 5.68 m, whereas ADC achieves 1.00 m. Although these traversals are not a controlled weather intervention, the results support useful cross-session/cross-weather generalization within the evaluated route.

  2. Meaningful controlled-corruption evidence. Models are frozen and evaluated on identical frames and reference poses. Under Snow-L5, ADC’s median translation error increases from 0.86 to 1.06 m (+23.3%), versus +155.4% for TempLoc and +277.0% for LightLoc. ADC also has the smallest relative mean rotation-error increase at Snow-L5. These positive results should be considered alongside the mean translation-error behavior.

  3. Complementary component gains and potentially practical cost. In Table 5, APFR reduces Boreas rotation error from 1.96° to 1.26° on top of H&P; with EdgeLoss already present, it reduces translation error from 1.13 to 0.98 m. The reported 72 MB model and 60 ms latency are promising, subject to clarification of the timing boundary. The appendices explicitly disclose traversal confounders and the selection of the illustrative segment.

Critical weaknesses

1. The incremental contribution over the closest prior designs is insufficiently established

The position clustering, regional prediction, normalization, perturbation, and conditioning in Appendix B.1 are closely related to LightLoc’s Sample Classification Guidance (SCG). ADC introduces differences, including multi-rotation global aggregation, but should identify what is inherited and demonstrate the benefit of the changes over the original guidance mechanism. LightLoc, §3.1

Similarly, SGLoc’s Inter-Geometric Consistency Constraint (IGCC) already supplements coordinate supervision by matching predicted and target pairwise distances. ADC uses a local radius graph and Euclidean distances, whereas the original uses L1 distances and sums over all pairs. These differences deserve analysis and direct comparison. SGLoc, §3.1, Eqs. 4–6

An effective combination can constitute a contribution. The current manuscript does not yet establish its added value with sufficiently specific comparisons to these neighboring designs.

2. The ablations do not isolate regional conditioning from rotation aggregation

Table 5 bundles sparse encoding, rotation aggregation, and regional conditioning into H&P. As formulated in Eqs. 4–6, the multi-rotation global descriptor conditions regression through the 25-dimensional regional response. Table 6 shows a large Boreas improvement from one to four views (5.09 to 0.98 m), compared with a much smaller Oxford change (0.62 to 0.59 m). The relationship between the single-view configuration and the 2.50 m vanilla model in Table 5 also needs an explicit account of their architectures and training protocols.

These results make regional-prior reliability a useful diagnostic for understanding the multi-view gain. Per-traversal region-classification accuracy, a matched no-prior ablation, and an oracle-region diagnostic could help separate the contribution of regional disambiguation from the other H&P changes. The oracle's conditioning representation and training setup should be specified, and reference-pose information should be used only for this diagnostic. Evaluating the prior on paired clean/corrupted inputs would further connect regional prediction to weather-related localization errors.

3. The results support robustness gains, but do not directly establish the claimed weather-handling mechanisms

ADC also achieves a substantial gain in clear weather: on the sunny test traversal (2021-03-02), ADC achieves 0.98 m translation error versus 3.19 m for TempLoc, a reduction of approximately 69.3%, similar to the margins on the rain and snow traversals. This indicates that the benefit extends to general localization accuracy. Additional evidence is needed to identify the contribution of the proposed weather-handling mechanisms, and this distinction should be reflected in Fig. 1 and Contribution 4. As Appendix A acknowledges, the real traversals also differ in traffic, session, and other scene conditions, limiting causal attribution to precipitation.

APFR is described as suppressing precipitation-related feature responses, but this effect is not directly measured. Likewise, if coordinate targets are generated as Xᵢ = Rpᵢ + t from observed points and reference poses, their Euclidean pairwise distances equal those of the observations. EdgeLoss then encourages a locally rigid mapping of observed geometry; it does not itself supply snow-free surface targets. It may nevertheless improve generalization.

The accumulation analyses transform the same input scans using different predicted poses. Sharper structures support better alignment, rather than independently identifying precipitation suppression. Component ablations on paired clean/corrupted frames, together with coordinate-error or correspondence measurements for known synthetic clutter and structural returns, would make the mechanism claims more convincing. Gate statistics could be supporting evidence, but would not alone establish causal suppression.

4. Synthetic-test coverage and error-distribution changes require further characterization

Appendix D describes a short window from one sunny traversal without frame indices, sample count, selection criteria, or random repetitions. For example, TempLoc’s clean error is 2.36 m in this window (Table 3), versus 3.19 m over the sunny traversal (Table 1). The window coverage and selection rule are needed to assess how representative the corruption results are.

Under Snow-L5, ADC’s relative mean translation-error increase exceeds TempLoc’s (+171.9% versus +154.7%), but its absolute increase is smaller (+1.65 versus +3.65 m), and its median and mean rotation behavior are more favorable. These statistics should be interpreted together.

The mild Snow-L1 setting is also informative: ADC’s mean translation error rises from 0.96 to 1.67 m (+74.0%), while its median moves only from 0.86 to 0.90 m. TempLoc’s mean increases from 2.36 to 3.32 m (+40.7%). ADC thus has the larger percentage increase but the smaller absolute increase (+0.71 versus +0.96 m). This motivates investigating which frames account for the additional error; it does not establish a higher failure rate or a worse tail than the baseline.

CDFs, high quantiles, threshold-based failures, and paired per-frame changes across prespecified windows or a larger sequence, with repeated corruption seeds, would clarify reliability. Relating these cases to regional-prediction errors, heading, or scene geometry would connect this analysis to the mechanism questions above.

5. Important training and comparison details remain unclear

Appendix B.1 gives the source of regional labels but not the classification loss or training stages. A separately pretrained classifier need not have its loss included in Eq. 14, but the procedure must be specified. The validation/model-selection protocol, encoder initialization, retained components, and training budgets for Tables 5 and 6 should also be documented. Repeated runs are particularly important for interpreting small ablation differences.

The paper states that splits and preprocessing are shared, but the Boreas training or adaptation procedures for LightLoc, RALoc, and GTRLoc should also be documented. TempLoc uses temporal correspondence and fusion, so its native inputs, sampling interval, and initialization at short-window boundaries should be specified. TransLoc4D was designed for 4D radar; its LiDAR adaptation and subsequent registration pipeline should be explained. TempLoc, TransLoc4D

Questions for authors

The following clarifications and targeted evidence would most affect my assessment:

  1. Closest-prior comparisons: What precisely changes relative to SCG and IGCC? Under a fixed encoder, reasonable matched training budgets, and the same pose solver, how do the original designs compare with ADC’s variants? Can rotation aggregation and regional conditioning within H&P be separated?

  2. Regional-prior diagnosis: What are the region-classification accuracies for the one-, two-, three-, and four-view configurations, by traversal and under paired synthetic corruptions? How do a matched no-prior ablation and a clearly specified oracle-region diagnostic perform? Does this help explain the 5.09 m single-view result relative to vanilla?

  3. Weather mechanisms and error distribution: Are coordinate targets derived entirely from observed scans and reference poses, or are clean-surface targets/static-point filters used? Do APFR and EdgeLoss reduce additional error under paired corruption? Please identify the synthetic window and settings, and characterize per-frame changes and tail/failure statistics, especially at Snow-L1 and Snow-L5, across a broader prespecified range and repeated corruptions.

  4. Reproducibility and baseline adaptation: Please specify classifier training, model selection, and the exact Table 5/6 configurations. How were LightLoc, RALoc, and GTRLoc trained or adapted on Boreas, including voxel size, initialization/pretraining, epochs, and model-specific settings such as regional clustering where applicable? GTRLoc’s mean rotation error differs substantially between Boreas and Oxford (7.40° versus 1.03°), making the adaptation details particularly useful. How are TempLoc’s temporal inputs/window initialization and TransLoc4D’s modality/registration handled? Are the main ablation effects stable across training seeds?

Reviewer_HKebOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes ADC, a scene coordinate regression (SCR) method for map-free LiDAR localization that is claimed to be robust to rain and snow. The architecture has three parts: a sparse convolutional encoder whose features are aggregated over a group of four 90-degree rotations of the input to form a global descriptor, which is then mapped to a 25-dim regional prior and concatenated to per-point features; a per-point squeeze-and-excitation style gating module (APFR); and an auxiliary loss (EdgeLoss) that matches pairwise distances between predicted scene coordinates and ground truth within 10m neighborhoods. Pose is solved with RANSAC as usual. Experiments are on Boreas (train on 4 sunny traversals of the Glen Shields route, test on 3 snow, 1 sun, 1 rain traversal) and on Oxford RobotCar. Reported mean translation error is 0.98m on Boreas vs 3.29m for the best SCR baseline (TempLoc) and 0.59m vs 0.72m on Oxford. There is a synthetic corruption study on a held-out sunny sequence, runtime/storage comparison, module ablations, a rotation-count ablation and a loss-weight sweep.

Critical strengths
  1. The problem is real and underexplored. Map-free LiDAR localization papers almost always evaluate on clear weather, and setting up a clear-train / adverse-test protocol on Boreas is a sensible contribution by itself. The ablations are reasonably complete for the proposed pieces (module ablation on both datasets, rotation count, lambda sweep), and the synthetic corruption study with frozen models on identical frames is the right kind of controlled experiment to include.

  2. Results are strong in absolute terms. Sub-meter error on all five Boreas test traversals with a 72MB model and no map, and a consistent (if modest) improvement on Oxford, would be a useful practical result if the comparisons hold up.

Critical weaknesses
  1. The paper’s own numbers do not support the weather-robustness story, which is the entire framing. In Table 1 the baselines are roughly equally bad on the sunny test traversal as on the snow ones: TempLoc is 3.17/3.15 on snow and 3.19 on sun, GTRLoc is 7.4 to 7.7 everywhere. So the 70% headline gap over TempLoc is there on clear weather too, it is not a weather effect, it is a dataset-level gap. The experiment that actually isolates weather is the synthetic study in Sec 4.3, and there ADC’s relative degradation is comparable to or worse than TempLoc’s (TE +63.5% vs +53.4% under Rain-L5, +171.9% vs +154.7% under Snow-L5). The appendix (A, last paragraph) essentially concedes that the real-weather results measure cross-session generalization on a known route rather than the effect of precipitation. What the evidence supports is “we built a more accurate SCR model whose advantage carries over to adverse conditions”, not “a framework specifically designed for adverse weather”. This requires rewriting the paper, not a rebuttal table.

  2. The technical novelty is thin, and nothing in the method is actually weather specific. All three modules are standard components under new names. Aggregating features over 90-degree rotations of the input is test-time augmentation / rotation-invariant pooling, long used in LiDAR place recognition (and RALoc, compared in the same table, is built around rotation awareness). APFR in Eq. 7-8 is exactly a squeeze-and-excitation gate applied per point; the paper’s claimed distinction from SE is just dropping the pooling step. EdgeLoss, penalizing pairwise distance discrepancies between predicted and ground-truth coordinates, is a common relative-distance regularizer in coordinate regression and registration. The regional classification prior resembles what LightLoc already does. None of these mechanisms sees weather labels, precipitation statistics, or any physical noise model, so the connection to rain and snow is asserted in the writing rather than built into the method. A combination of known pieces can still be a fine paper if honestly framed and rigorously compared, but here the framing claims a first-of-its-kind weather-robust design that the components do not deliver.

  3. The comparison on Boreas is likely unfair to the baselines, which undermines the main table. GTRLoc gets 0.75m/1.03° on Oxford but 7.67m/7.40° on Boreas, RALoc similarly collapses, and rotation errors above 7 degrees for a published NeurIPS method on a standard driving route look like a broken retraining rather than a property of the method. The paper only says all methods share identical splits and preprocessing, with no information on per-method tuning (voxel size, learning rates, augmentation) for the 128-beam Boreas data. Telling in this regard: the authors’ own vanilla baseline in Table 5, with none of the proposed modules, already reaches 2.50m on Boreas, beating every published SCR baseline in Table 1 by a large margin. So most of the reported gap comes from the backbone and preprocessing choices, not from the contributions, and the contributed modules account for 2.50 to 0.98. The obvious cheap alternative is also missing: run TempLoc or LightLoc with a standard de-snowing filter (DROR / LIOR style) on the input. Without that one cannot tell whether the architectural machinery is needed at all.

Questions for authors

Please report, for each baseline on Boreas, what tuning was done (voxel size, lr, epochs, augmentations) and whether the authors’ retrained numbers match any published Boreas results. Why does GTRLoc show 7+ degree rotation error there while being fine on Oxford?

The vanilla model in Table 5 (2.50m) already outperforms all SCR baselines in Table 1. What exactly is in “vanilla”, and how do you reconcile this with attributing the improvement to the three proposed modules?

Can you add a baseline of an existing SCR method combined with a classical snow/rain outlier removal filter on the input cloud, under the same protocol?

In Table 3 ADC’s relative degradation under Rain-L5 and Snow-L5 is not smaller than TempLoc’s. Do you agree that the real-weather gap in Table 1 therefore reflects a general accuracy advantage rather than weather robustness, and if not, what experiment isolates the weather effect in your favor?

The 4-rotation aggregation presumably needs four encoder passes. Is the 60ms latency with all four, and what is the single-view latency?

Is the training/test route identical (Glen Shields) for Boreas? If so please state clearly that evaluation is same-route relocalization, and consider a second route.

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

Related papers

Discussion (0)

Sign in to comment.