acceptodds
Under review as a conference paper at ICLR 2027

Faster, Fewer, & Farther: Granularity-Aligned Hierarchical LiDAR Relocalization in a Shared Spectral Geometry

Abstract

Global LiDAR relocalization requires transferable association across growing maps, where geographically distant places can share similar structure. Exhaustive flat frame retrieval preserves a simple comparison interface but incurs map-linear search and increasing exposure to distant aliases. Hierarchical retrieval reduces this search, yet makes coarse routing an irreversible geographic hypothesis-pruning operation: a discarded location cannot be recovered by fine matching. We identify spatial, granularity, and embedding-consistency mismatches: coarse candidates can be appearance-coherent yet geographically distant, routing may compare evidence at unequal granularities, and separately fitted stage representations may disagree under domain shift. F3Loc aligns support and granularity in a simple hierarchy with no routing-specific representation learning. Query context routes to spatially coherent map regions, after which the current frame localizes among the retained frames. The proposed shared-basis spectral calibration establishes a fixed comparison geometry before aggregation, and granularity-aligned routing preserves the resulting frame-neighborhood semantics from regional routing to final localization. Across Oxford, NCLT, and MCD, F3Loc discards 96.1–99.5% of map frames, yielding 2.35–8.90× speedups over same-encoder exhaustive retrieval. Mean translation error falls by 54–56% on Oxford and 31–52% on NCLT and MCD. Under the same no-target-training protocol, F3Loc achieves substantially lower localization error than industrial ANN and hierarchical retrieval controls, while attaining translation accuracy competitive with, and on several routes better than, in-domain regressors. Code will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
LiDAR Relocalization, Hierarchical Retrieval, Granularity alignment, Shared Spectral Geometry

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

F3Loc is a target-free LiDAR relocalization method that does not require training on the target scene and adopts a coarse-to-fine strategy. It applies a source-fitted spectral transform to frame descriptors, aggregating query context and contiguous map regions within a shared space, and subsequently performs frame matching within the selected region. F3Loc-Fast directly outputs the stored map poses, whereas GeoRefine adds a local registration step. The main contribution of this work lies in proposing a simple retrieval scheme that effectively reduces search overhead and mitigates false matches at long distances. The paper evaluates the method on the Oxford, NCLT, and MCD datasets, covering causal retrieval and two different descriptor encoders.

Critical strengths

The paper presents a clear and useful design for target-free LiDAR relocalization. F3Loc changes the support and granularity of the evidence used for retrieval while keeping one source-fitted comparison geometry across routing and final matching. This directly addresses the risk that coarse-to-fine routing discards the correct geographic hypothesis before fine matching. The work also demonstrates a degree of engineering innovation through the integration of these design choices into a practical and coherent relocalization pipeline.

Critical weaknesses

1.Although the paper claims that coarse routing is built upon "spatially coherent support", the actual algorithm (Section 3.3) relies entirely on the one-dimensional continuous frame indices of a single ordered mapping sequence, using a sliding window with a core of 50 frames plus 10 frames on each side to segment regions. This design works for maps from a single continuous capture, yet it suffers notable limitations in practical applications when confronted with multi-session maps, multi-pass trajectories, or unordered submaps.

2.Table 2 reports ANN lookup on CPU and F3Loc routing and matching on GPU, so the reported latency comparison is not hardware-matched. The accuracy and far-error results remain informative, and the comparison with exhaustive retrieval supports F3Loc’s efficiency. However, the main text does not establish the latency trade-off against HNSW+G under comparable hardware conditions.

Questions for authors
  1. Section 3.3 mentions spatially local graph partitions as an alternative to contiguous frame intervals. Could the authors clarify what properties such a partition should satisfy for F3Loc, and whether the method requires a single ordered traversal in practice? This would help establish how directly the current approach extends to multi-session maps or maps assembled from unordered submaps.

  2. Could the authors clarify how the latency comparison with HNSW+G should be interpreted given that the ANN lookup runs on CPU while F3Loc runs on GPU? If available, latency measurements under a matched hardware setup would help assess the practical trade-off.

3 comments
AuthorsOpenReview ↗

Response to Reviewer RUqg (1/2)

Thank you for your careful review and for recognizing the useful retrieval design and geographic-hypothesis perspective. We tested your map-construction and timing concerns directly. This part answers W1/Q1; Part 2 addresses W2/Q2. We have uploaded a revised manuscript, with changes highlighted.

R3.1 [W1/Q1]: Does map construction require a single ordered traversal?

The map need not come from one continuous traversal. Within-traversal groups consecutive frames from the same recording; map-pose groups nearby map positions across recordings, so observations from different dates can share a region. Both use cores of at most 50 frames with boundary support, retain every map frame, and need no query pose.

Only map grouping changes: regional scores aggregate calibrated frame descriptors, followed by current-frame matching within selected regions. Storage order is irrelevant if region membership is preserved. Causal (C) uses current/past query frames; symmetric (S) also uses future frames. Neither changes map grouping or current-frame fine matching.

Table R1. Maps from multiple NCLT traversals; mean TE (m) and failures >50 m (FAR).

Map groupingQuery context2-traversal TEFAR3-traversal TEFAR
Raw flat, no G; no regionsCurrent frame2.6541232.451104
Flat + G; no regionsCurrent frame1.816621.74553
Within-traversal: same-recording time neighborsCurrent + past (causal)1.559341.47923
Within-traversal: same-recording time neighborsPast + current + future (symmetric)1.462131.46315
Map-pose: spatial neighbors across recordingsCurrent + past (causal)1.475261.3898
Map-pose: spatial neighbors across recordingsPast + current + future (symmetric)1.416121.3667

The first map combines the 02-18 and 03-17 recordings: 32,073 frames and 25,261 eligible 02-12 queries. Adding 3,507 available 06-15 frames gives 35,580 map frames and 25,361 eligible queries. All rows use the same map-specific query mask; encoder/G are unchanged and the 02-12 query recording is absent from the map. Within-traversal C/S pools contain 221/218 and 222/218 frames; map-pose pools contain 153 and 169.

Both constructions improve TE and FAR over Flat + G. A spatial graph partition should cover all searchable map frames, group geographically neighboring observations, keep region sizes bounded, and retain boundary overlap so adjacent regions do not lose useful candidates. The map-pose construction meets these conditions without a global frame order. The experiments test overlapping multi-session maps; arbitrary road-junction graph topology remains untested.

R3.2 [W1/Q1, continued]: What happens when map order or coverage changes?

We shuffled 200-frame submaps and rebuilt regions within each fragment. This tests reordered local submaps, not arbitrary graph partitions. We also removed frames in the central 10% of map X extent and split regions at the gaps. Query context and W=40/50, Top-4, encoder/G stay fixed.

Table R2. Mean TE (m), identical edited maps and query masks. C/S: causal/offline symmetric.

MapRouteRaw flatFlat + GF³Loc CF³Loc S
ShuffledOxford 17-138.9733.7084.0542.161
ShuffledNCLT 02-122.5911.8031.5451.523
ShuffledMCD day1012.61612.1598.3195.907
GapOxford 17-139.3623.9513.6232.249
GapNCLT 02-122.3181.8031.5931.502
GapMCD day1012.21911.9537.6696.339

Gap-test eligible/total queries are 38,461/42,778, 23,151/25,546, and 2,394/3,231. These tests concern spatially missing map sections, distinct from combining recordings. Evaluation covers locations still represented in the map. Local neighborhoods and boundary overlap remain necessary; arbitrary index adjacency is insufficient.

AuthorsOpenReview ↗

Response to Reviewer RUqg (2/2)

R3.3 [W2/Q2]: How do F³Loc and HNSW+G compare on matched hardware?

We apologize for the earlier GPU-lookup label: both HNSW and F³Loc retrieve on CPU. The revised manuscript reports the full retrieval-baseline timing comparison in Table 13.

Table R3. Updated matched timing (ms/query), calibrated descriptors; Lookup and Full-query are separated.

MapBExhaustive+G lookupHNSW+G lookupF³Loc lookupHNSW+G full-queryF³Loc full-query
Oxford1413.71315.81313.16034.23132.250
Oxford16104.4334.4954.9446.4276.899
NCLT1273.97615.9899.58134.11428.681
NCLT1668.7964.5534.0006.6096.147
MCD161.30713.2024.31931.69923.393
MCD1615.4633.7002.7335.7154.725

We use 128 identical queries/map, four threads on a Xeon Gold 6326, RTX 4090 encoding, and one warm-up with three interleaved passes. HNSW reuses its saved index (M=32, efConstruction=200, efSearch=128); F³Loc uses causal W=40, Top-4, 50-frame cores, and 10-frame boundaries. Lookup covers CPU candidate selection, fine matching and MapPose; Full-query adds encoding, calibration, and online transfers. Offline index construction is excluded. B=16 is amortized ms/query, not single-query response latency.

On MCD, F³Loc improves route-wide TE (7.22 vs. 12.16 m) and matched B=16 full-query time (4.725 vs. 5.715 ms/query) over HNSW+G. On Oxford, HNSW+G is slightly faster (6.427 vs. 6.899 ms/query). Section 4.3 states this trade-off, while revised Table 13 gives matched B=1/B=16 CPU lookup and full-query times for the remaining retrieval baselines in Table 2 (including the original single-cell IVF-Flat). Sparse routing gives approximately 15–32× faster CPU lookup than matched dense temporal search, not an end-to-end speedup. Earlier per-system times followed a separate protocol.

Table R5. All eight routes; mean TE (m), same calibrated banks.

RouteHNSW + GF³Loc causalF³Loc symmetric
Oxford 15-135.014.785.53
Oxford 17-134.693.452.15
Oxford 17-144.543.093.11
NCLT 02-121.811.511.51
NCLT 02-191.991.831.42
NCLT 03-312.001.531.46
NCLT 05-266.385.524.99
MCD day1012.167.226.04
Route macro4.823.623.28

Oxford 17-13's 3.45 m is recomputed from archived MapPose IDs (submitted summary: 3.43 m). The submitted HNSW+G reference is 4.26 m; the rebuilt-index evaluation gives 4.69 m.

Causal F³Loc lowers TE on all eight routes versus HNSW + G. Symmetric uses future context and is offline. Against HNSW+G, F³Loc has lower CPU lookup in five of six batch/map settings, but HNSW+G is faster in Oxford B=16 full-query timing. This does not alter the submitted accuracy comparison.

The earlier ANN timing used a different query sample and runtime protocol. The new comparison fixes queries, hardware, indices, and timing scopes.

Please let us know if any questions or concerns remain; we will do our best to address them.

I appreciate the author’s detailed response and experiments, particularly the clarifications and corrections regarding the CPU and GPU. In my view, the author’s response has addressed the issues I raised. I will maintain my score of 3.

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

The paper studies hierarchical LiDAR relocalization, focusing on the risk of discarding the correct geographic region during coarse retrieval. The authors identify three sources of mismatch: spatially distant but visually similar regions, inconsistent evidence granularity between routing and final matching, and changes in descriptor geometry across retrieval stages. F3Loc addresses these issues by partitioning an ordered map traversal into overlapping regions and using temporal query context to select a small set of candidates. A source-fitted spectral transformation is applied to frame descriptors before aggregation, so coarse routing and final frame matching use the same calibrated similarity space. The approach does not require target-scene training. Experiments on Oxford, NCLT, and MCD show substantial reductions in candidate comparisons and mean translation error relative to exhaustive retrieval. The results are particularly strong in reducing large localization errors caused by geographically distant matches.

Critical strengths
  1. Clear problem framing supported by evidence. The paper frames coarse routing as irreversible pruning of geographic hypotheses and substantiates each of the three identified mismatches through targeted experiments. For example, frame-to-region routing without granularity alignment reduces the search space but increases localization error (Table 3). Similarly, independently learned stage mappings substantially degrade performance even with calibrated inputs (Table 2).

  2. Systematic ablation studies. The authors systematically isolate the contributions of individual components through incremental ablations, all combinations of coarse- and fine-stage geometries, stage-wise failure analysis, transfer experiments with EgoNN, and sensitivity analyses. Candidate accounting and execution protocols are also documented in sufficient detail to support reproducibility.

  3. Strong performance gains. F3Loc maintains a small, nearly constant candidate pool as map size increases, while substantially reducing localization error, primarily by suppressing geographically distant failures. These improvements persist even against HNSW using the same geometry. When combined with GeoRefine, F3Loc also outperforms in-domain regressors across all NCLT routes without requiring target-scene training.

Critical weaknesses
  1. The contribution of hierarchical routing is not sufficiently isolated from spectral calibration. Table 3 shows that applying G to exhaustive retrieval alone reduces mean TE from 8.97 to 3.71 m on Oxford 17-13, compared with 3.43 m for causal F3Loc and 2.15 m for the symmetric variant. This raises questions about how much of the improvement comes from spectral calibration rather than the hierarchy itself. Although HNSW+G and calibration variants are partially evaluated, matched comparisons across all routes are missing. Moreover, Eq. 11 follows from the inner-product structure of aggregated descriptors rather than a property unique to the proposed method.

  2. Baseline fairness and generalization remain insufficiently established. F3Loc uses 41–101 query frames, while most baselines rely on a single frame. The weak performance of the sequence-based baselines makes it difficult to isolate the benefit of region partitioning from temporal context. A partition-free temporal baseline using the same G and window would strengthen the comparison. The method also assumes an ordered map traversal and has not been tested on branched maps, reverse traversals, or varying speeds. In addition, it is unclear whether hyperparameters were selected exclusively from source-domain validation, since Figs. 8 and 13 present target-route results. Finally, median and percentile TE would help assess whether the reported mean improvements extend beyond a small number of large localization errors.

Questions for authors
  1. Contribution of G versus hierarchical routing. Could the authors provide matched comparisons of exhaustive retrieval with G, conventional sample-weighted PCA whitening, HNSW+G, and causal/symmetric F3Loc across all evaluated routes? Table 3 shows a large improvement from G alone on Oxford 17-13, while the remaining gain differs substantially between causal and symmetric routing. I would like to understand which part of the improvement is attributable to the hierarchy itself.

  2. Hyperparameter selection. Were W, Top-K, region length, source-domain mixture, and the choice between LoRA and full fine-tuning finalized using only source-domain validation? In particular, were the target-route results in Figs. 8 and 13 used only for post-hoc analysis, or did they influence any configuration choices?

  3. Temporal context and baseline comparison. How would F3Loc compare with a partition-free sliding-window or SeqSLAM-style baseline using the same G and temporal window? Also, could the authors clarify how the sequence-based baselines were tuned and whether the exhaustive retrieval timing uses an optimized GPU implementation under comparable batch and hardware conditions?

  4. Error distribution and routing failures. Could the authors report median and higher-percentile TE in addition to mean TE and FAR? It would also be useful to report the beneficial inversion rate alongside the 10.54% harmful inversion rate discussed in Appendix C. These results would help assess whether the proposed geometry consistently preserves useful nearest-neighbor relationships and how broadly the localization improvements are distributed.

4 comments
AuthorsOpenReview ↗

Response to Reviewer gnME (1/3)

Thank you for recognizing our problem framing and ablations. Part 1 examines G versus routing and parameter selection; Part 2 tests matched-context retrieval and timing; Part 3 covers error distributions and map/speed robustness. We have uploaded a revised manuscript, with changes highlighted.

R2.1 [W1/Q1]: What comes from G versus routing, and what does Eq. 11 establish?

With G fixed, temporal regional retrieval lowers macro TE by 20.1%/27.6% and removes 96.1–99.5% of final frame comparisons. Encoder, map, queries, and 3D metric are matched. C/S: causal/offline symmetric.

Table R1. Mean TE (m), all eight routes.

RouteFlat + GHNSW + GF³Loc CF³Loc S
Oxford 15-134.795.014.785.53
Oxford 17-133.714.693.452.15
Oxford 17-143.854.543.093.11
NCLT 02-121.801.811.511.51
NCLT 02-191.981.991.831.42
NCLT 03-311.992.001.531.46
NCLT 05-265.976.385.524.99
MCD day1012.1612.167.226.04
Macro4.534.823.623.28
Reduction vs Flat + GReference20.1%27.6%

Oxford 17-13's 3.45 m is an archived-MapPose rerun (submitted: 3.43 m). Rebuilt HNSW+G gives 4.69 m (submitted: 4.26 m; M=32, efConstruction=200, efSearch=128).

Macros weight routes equally. Raw flat/C/S is 6.69/5.94/5.32 m: routing helps before G, though raw routing can discard positives. On Oxford, G raises coarse pool coverage 98.61%→99.85% (+524 queries).

Appendix J (Table 17) reports the fixed EgoNN checkpoint: on Oxford 17-13, raw flat/temporal hierarchy/hierarchy + G TE is 22.80/17.93/15.14 m; on NCLT 02-12, it is 15.53/9.00/7.11 m. These results show that hierarchical retrieval is effective with EgoNN.

Eq. 11 expresses aggregation of the same calibrated frame kernel, not a new theorem or a guarantee of rank preservation. Oxford coarse/fine swaps (manuscript Table 16) give raw/G 12.12 m versus G/G 3.43 m, illustrating the impact of pruning before fine matching.

Table R1a. Conventional sample-weighted pooled PCA whitening; eight-route macro TE (m), matched dimensions/normalization.

CalibrationFlatHierarchy CHierarchy S
Sample-weighted pooled PCA4.543.913.55
Shared G4.533.623.28

Flat results are nearly tied; G performs better under hierarchical routing. The sample-weighted pooled-covariance control is defined in revised Appendix B.1 and shown in Fig. 5(c).

R2.2 [W2/Q2]: How were parameters selected, and how sensitive are the gains to routing settings?

Encoder/G fitting uses source data. Routing settings (C/S W=40/50, Top-4, 50-frame cores), source mixture, and LoRA versus full tuning were informed by NCLT 02-12/03-31 development routes; Oxford/MCD reuse those choices. We report NCLT sensitivity and a separate, post-submission Sejong check that selects W/core at fixed Top-K=4, not a replacement history for the submitted settings.

Table R1b. Routing sensitivity on NCLT 02-12.

ParameterTested valuesMean TE range (m)
Symmetric radius W10, 20, 30, 40, 501.51–1.63
Top-K2, 4, 6, 81.39–1.68
Region core40, 50, 601.48–1.62

Top-K=2 gives 1.39 m on NCLT 02-12, but Top-K=4 retains more geographic candidates. We apply the same K to every evaluation route rather than choosing route-specific optima.

For the independent Sejong check, Sejong01 maps Sejong02 (20,143 eligible queries), separate from all eight evaluation routes. Keeping encoder/G and Top-K=4 fixed, we chose W/core by Sejong validation TE, then evaluated the choices on the eight routes without further tuning.

Table R1c. Submitted versus validation-selected configurations, with Top-K fixed at 4. Mean TE is macro-averaged equally across eight target routes.

ConfigurationContextWTop-KRegion coreMean TE (m)
Raw flatSingle frame–––6.69
Flat + GSingle frame–––4.53
Submitted F³LocCausal404503.62
Validation-selectedCausal504503.67
Submitted F³LocSymmetric504503.28
Validation-selectedSymmetric504603.41

On Sejong, submitted C/S settings are within 0.43%/1.86% of the best Top-4 choices. Validation-selected settings yield 3.67/3.41 m eight-route macro TE (Flat + G: 4.53 m).

AuthorsOpenReview ↗

Response to Reviewer gnME (2/3)

R2.3 [W2/Q3]: Do matched temporal controls isolate the effect of region partitioning?

We added matched-context IVF and SeqSLAM-style controls in the revised manuscript (Appendix F.1, Table 14), using G, the same query frames (up to 40 past for causal, 50 on either side for offline symmetric), and current-frame fine matching. The dense–sparse control in Table 15 then isolates the cost of region partitioning.

Table R2. Mean TE (m), matched query context.

MethodContextOxfordNCLTMCD
IVFCausal8.002.0814.69
SeqSLAM-styleCausal18.324.7220.15
F³LocCausal3.451.517.22
IVFSymmetric5.411.7013.08
SeqSLAM-styleSymmetric25.222.2713.76
F³LocSymmetric2.151.516.04

Routes: Oxford 17-13/NCLT 02-12/MCD day10. In the revised manuscript, Table 2 reports single-cell, current-frame IVF-Flat (33.82 m on Oxford). Our matched-context IVF probes four lists to parallel F³Loc's Top-4 regions, with G, the same query window, and current-frame fine matching (8.00 m causal; 216/222 Oxford candidates). Appendix Table 13 times the original single-cell IVF-Flat setting, not the four-list matched-context variant.

Sequence-baseline settings. The SeqSLAM-style and SeqLPD-style baselines were adapted to the same frozen BEVPlace++ descriptors, with implementation-specific settings fixed for each control. The additional SeqSLAM-style comparison uses the same temporal windows as F³Loc and a Top-220 candidate budget, comparable to its fine-search budget. For SeqNet-style, the LiDAR heads were trained on source descriptors, with checkpoints selected by source-training loss. SeqOT (reply in JJmF) was evaluated using its official NCLT-pretrained model without retraining on Oxford.

The dense ablation searches 41-frame windows at each map position and expands 220 frames; sparse routing selects Top-4 regions with the same current-frame fine matcher.

Table R2a. Complete causal ablation; mean TE (m).

RouteDense + GF³Loc
Oxford 15-135.184.78
Oxford 17-132.793.45
Oxford 17-143.213.09
NCLT 02-121.561.51
NCLT 02-191.711.83
NCLT 03-311.581.53
NCLT 05-265.775.52
MCD day106.907.22
Macro3.593.62

Under the same G and temporal context, dense and sparse retrieval have nearly identical macro TE (3.59 and 3.62 m). F³Loc achieves this accuracy with about 40× fewer coarse keys. Both retain the frame bank.

Table R2b. Matched causal B=1 CPU lookup (ms/query), four threads, 128 identical queries/map.

MapDense temporal+GF³Loc
Oxford416.89113.160
NCLT277.6519.581
MCD64.0714.319

This matched comparison yields approximately 15–32× faster CPU lookup, not end-to-end speedup, with virtually identical macro TE (3.59/3.62 m). Context, calibrated frame bank, and fine matcher are identical. Appendix Table 13 reports both timings.

R2.3 [W2/Q3, continued]: Were search hardware, batching, and timing scopes matched?

We apologize for the earlier GPU-lookup label: HNSW, F³Loc, and the exact exhaustive lookup in this test all run on CPU. We used 128 identical queries/map, four Xeon Gold 6326 threads, RTX 4090 encoding, and three post-warm-up passes. These are matched CPU comparisons.

Table R3. Updated matched ms/query, with calibrated descriptors; the revised manuscript's Appendix Table 13 additionally lists all Table 2 retrieval baselines.

MapBExhaustive+G lookupHNSW+G lookupF³Loc lookupHNSW+G full-queryF³Loc full-query
Oxford1413.71315.81313.16034.23132.250
Oxford16104.4334.4954.9446.4276.899
NCLT1273.97615.9899.58134.11428.681
NCLT1668.7964.5534.0006.6096.147
MCD161.30713.2024.31931.69923.393
MCD1615.4633.7002.7335.7154.725

Lookup covers CPU search and MapPose; full-query adds encoding, calibration, transfers, and online updates. Offline indexing is excluded. B=1 is latency; B=16 is amortized ms/query.

On MCD, F³Loc improves route-wide TE (7.22 vs. 12.16 m) and matched B=16 full-query time (4.725 vs. 5.715 ms/query) over HNSW+G. On Oxford, HNSW+G is slightly faster (6.427 vs. 6.899 ms/query). Original Table 1 timings use a separate protocol.

The 476.4 ms result used another protocol. In a separate Oxford B=1 Flat+G profile, CPU lookup takes 412.527 ms and GPU encoding 19.218 ms (432.583 ms total). See Appendix Table 13 for matched timing.

AuthorsOpenReview ↗

Response to Reviewer gnME (3/3)

R2.4 [W2/Q4]: Do mean gains reflect tail suppression, and are beneficial neighborhood changes counted?

Table R4. TE (m) and FAR (>50 m), Oxford 17-13 and MCD day10. Raw flat uses uncalibrated descriptors.

RouteMethodMeanMedianP95P99FAR
Oxford 17-13Raw flat8.9731.9604.358359.224669
Oxford 17-13F³Loc Causal3.4531.9304.0735.14984
Oxford 17-13F³Loc Symmetric2.1521.9264.0635.08723
MCD day10Raw flat12.6161.37874.906261.781160
MCD day10F³Loc Causal7.2201.25937.954116.019109
MCD day10F³Loc Symmetric6.0411.22723.247133.75778

Median errors remain similar, while severe long-distance localization failures are greatly reduced. On Oxford, P99 falls from 359.224 m to 5.149/5.087 m (C/S), and FAR drops from 669 to 84/23. MCD shows the same pattern in P95 (74.906 to 37.954/23.247 m) and FAR (160 to 109/78). Thus, the main improvement comes from suppressing catastrophic mismatches rather than reducing typical localization errors.

Table R4a. Learned S1 mapping; 42,762 queries, original 5 m planar (XY) positive mask.

ChangeQueries% of all% of initial group
Correct → wrong4,39710.282510.54
Wrong → correct1120.261910.87

Initially, 41,732 queries have correct positive–negative orderings and 1,030 do not. After mapping, 37,335 stay correct and 918 stay wrong. The ordering correct rate falls from 97.59% to 87.57%. Both inversion rates now follow Appendix C's original XY margin protocol, rather than localization recall. This inversion audit concerns the source-trained S1 stage mapping, not G itself; G's coarse/fine impact is tested by the geometry swaps and fixed-pool ranking controls.

R2.5 [W2]: What about branched maps, reverse traversals, and varying speeds?

Branched-map evidence and boundary. A three-traversal NCLT map (02-18/03-17/06-15) tests overlapping passes without one global sequence. Temporal regions stay within recordings; spatial/map-pose regions group nearby poses across recordings, using no query GT.

Table R5. Three-traversal map (35,580 frames, 25,361 queries), mean TE (m).

RetrievalTE
Raw flat / Flat+G2.451 / 1.745
Temporal regions C/S1.479 / 1.463
Map-pose regions C/S1.389 / 1.366

Both region types beat Flat+G. Removing the central 10% of map X extent and splitting at gaps gives Oxford/NCLT/MCD Flat+G→C/S TE of 3.951→3.623/2.249, 1.803→1.593/1.502, and 11.953→7.669/6.339 m on still-covered queries. These results support locally coherent partitions across sessions and gaps, not general branching road-graph topology or junction-specific robustness, which remain untested.

Varying speeds (measured, not simulated). We stratified 24,953 original NCLT 02-12 map-positive queries into GT-speed tertiles using timestamps and ~1 s planar displacement; no resampling or rerun. Query IDs are matched across methods.

Table R5a. Natural speed groups; mean TE (m).

Speed (m/s)NFlatFlat+GF³Loc CF³Loc S
Low (≤1.24)8,3183.482.441.741.73
Mid (1.24–1.41)8,3172.091.501.411.44
High (1.41–3.60)8,3182.201.461.371.37

F³Loc improves over flat retrieval across all three speed groups. Causal mean TE falls from 3.48 to 1.74 m at low speed, 2.09 to 1.41 m at medium speed, and 2.20 to 1.37 m at high speed; symmetric results show the same consistent improvement.

Simulated reversal. Rotating Oxford/NCLT query BEVs by 180° leaves TE (three decimals), R@1 and FAR unchanged for Flat+G and F³Loc C/S; it does not test actual reverse-direction traversals.

Please let us know if any questions remain.

My concerns are addressed. I will maintain my score.

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

This paper studies global LiDAR relocalization based on descriptor retrieval. It treats coarse-to-fine routing as irreversible hypothesis pruning and points out three kinds of mismatch in hierarchical retrieval: a spatial mismatch, where coarse candidates look alike but are far apart geographically; a granularity mismatch, where single frames are compared against region summaries; and an embedding-consistency mismatch, where mappings fitted separately for each stage change the nearest-neighbor structure in the target domain. F³Loc builds on a frozen encoder with source-only LoRA adaptation. It combines a shared spectral geometry, map regions built from contiguous trajectory spans, and granularity-aligned routing. Experiments cover Oxford, NCLT, and MCD.

Critical strengths

The paper frames coarse routing as irreversible pruning and measures it directly through candidate-pool coverage (Table 8) and a stage-wise failure breakdown (Table 10), which is a useful diagnostic. The method needs no routing-specific training, and the comparison is well controlled: all candidate-generation methods share the same frozen BEVPlace++ descriptors and query set (Tables 2 and 6). The 2×2 geometry swap (Table 8) and the fixed-pool ranking test (Table 5) separate the effect of G on coverage from its effect on ranking. Long-tail errors drop clearly in hard cases (MCD mean TE from 12.62 m to 7.22/6.04 m; NCLT 05-26 from 9.14 m to 5.52/4.99 m), and the same trend holds with a second encoder (EgoNN, Table 9).

Critical weaknesses
  1. Most of the gain comes from the whitening G, not from the hierarchy.

In Table 3, adding G to flat exhaustive retrieval alone lowers TE from 8.97/2.59 m to 3.71/1.80 m, while the hierarchy without G reaches only 8.04/1.93 m. The headline 2.15 m result uses the offline symmetric variant; the causal version reaches 3.43 m, only 0.28 m better than flat retrieval with G, and that gap mixes temporal context with region pruning. Controls such as flat retrieval or SeqSLAM/IVF-Flat with G and the same query context, and appearance-clustered versus trajectory regions under the same geometry, are needed to support the central claim.

  1. The comparisons are unfair and the speed claim is overstated.

F³Loc uses 41 to 101 query frames (including 50 future frames in the symmetric version), while most baselines use single frames, yet the offline variant is bolded alongside online methods and its 7.2 ms excludes the wait for future frames. The reimplemented sequence baselines all do worse than single-frame retrieval (18.74 to 44.80 m vs 8.97 m), and SeqOT and the official SeqNet are missing. The flat baseline's 476.4 ms (batch 1) is unusually slow for about 40,000 frames and needs a latency breakdown. The training data of the baseline checkpoints should also be stated.

  1. Hyperparameters appear to be tuned on test routes.

The "post-hoc sensitivity" sweeps in Appendix E run on test routes, and Top-4, 50-frame regions, and W=40 are justified by test-route TE (2.695 m is exactly the mean of the causal entries in Table 1). This risks test-set leakage.

  1. Novelty is limited.

The method combines sequence-averaged descriptors, trajectory chunking, coarse-to-fine Top-K, and a PCA-whitening variant, none of which is new. The advantage over standard PCA whitening appears only as points in Figure 5c, and whitening fitted on the unlabeled map descriptors, a natural baseline, is missing.

  1. Others

The caption of Figure 3 contains a leftover from revision. Its first sentence, "Note: Shared spectral geometry is applied first to frame descriptors.", is in blue and repeats what the rest of the caption says. The abstract and introduction pile up many concepts, such as "irreversible hypothesis-pruning" and "comparison geometry", which makes them hard to follow. Key evidence, including Table 8, the MCD results, and Appendix E, sits in Appendices A–L, and the full paper runs to 25 pages. I suggest moving the core evidence into the main text. Table 1 has unexplained gaps. HOTFormerLoc shows "–" for Oxford and "–/24.6" for latency with no explanation. The authors should also say whether the GeoRefine latency includes ICP.

Questions for authors

See above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer JJmF (1/3)

Thank you for recognizing the geographic-pruning analysis and the controlled ablations. We have uploaded a revised manuscript, with changes highlighted.

Main result. With G fixed, temporal regional retrieval lowers eight-route macro TE by 20.1%/27.6% (causal/symmetric). Dense retrieval isolates the accuracy benefit; sparse routing retains similar causal accuracy with 96.1–99.5% fewer final frame comparisons.

R1.1 [W1]: Does calibration explain the gains, and what does hierarchy add across routes and encoders?

G improves frame comparisons; temporal routing rejects distant alternatives. All entries share each route's encoder, map, queries, and 3D metric. C/S: causal/offline symmetric; flat: exhaustive search.

Table R1. Mean TE (m), all eight routes.

RouteRaw flatFlat + GRaw H, CRaw H, SF³Loc CF³Loc S
Oxford 15-137.584.795.996.674.785.53
Oxford 17-138.973.719.368.043.452.15
Oxford 17-148.083.858.928.803.093.11
NCLT 02-122.591.801.681.931.511.51
NCLT 02-192.041.981.291.331.831.42
NCLT 03-312.481.991.671.691.531.46
NCLT 05-269.145.979.557.865.524.99
MCD day1012.6212.169.046.217.226.04
Route macro6.694.535.945.323.623.28
Reduction vs Flat + GReference20.1%27.6%

Oxford 17-13's 3.45 m is recomputed from archived MapPose IDs (submitted summary: 3.43 m). The submitted HNSW+G reference is 4.26 m; the rebuilt-index evaluation gives 4.69 m, retaining M=32, efConstruction=200, and efSearch=128.

Macros weight routes equally. Coarse errors can discard positives. On Oxford 17-13, G raises positive coverage from 98.61% to 99.85%, keeping 524 additional queries reachable. On MCD, G alone lowers TE from 12.62 to 12.16 m; routing reaches 7.22/6.04 m.

Appendix J (Table 17) tests frozen EgoNN: raw flat/temporal hierarchy/hierarchy + G TE is 22.80/17.93/15.14 m on Oxford 17-13 and 15.53/9.00/7.11 m on NCLT 02-12. This demonstrates the effectiveness of hierarchical retrieval with EgoNN. Eight-route R@1@5m stays close: Flat + G 96.33%, F³Loc C/S 96.28%/96.45%, consistent with fewer severe geographic errors.

R1.2 [W1/W2]: Do matched temporal controls separate context from regional routing?

We added matched-context IVF and SeqSLAM-style controls (revised Appendix F.1, Table 14). They share G, the query windows (up to 40 past frames causal, 50 on either side offline symmetric), and current-frame fine matching.

Table R2. Mean TE (m), matched query context.

MethodContextOxford 17-13NCLT 02-12MCD day10
IVF, four listsCausal8.002.0814.69
SeqSLAM-styleCausal18.324.7220.15
F³LocCausal3.451.517.22
IVF, four listsSymmetric5.411.7013.08
SeqSLAM-styleSymmetric25.222.2713.76
F³LocSymmetric2.151.516.04

In the revised manuscript, Table 2 uses single-cell, current-frame IVF-Flat (33.82 m on Oxford). Matched-context IVF instead probes four lists to parallel Top-4 regional routing, with G, the shared query window, and current-frame fine matching (8.00 m causal); Oxford IVF/F³Loc pools contain 216/222 causal and 207/216 symmetric frames. Revised-manuscript Appendix Table 13 times the single-cell baseline. SeqSLAM-style scores paths with local contrast, both directions, velocities 0.8–1.2, and Top-220 reranking.

At Top-4, appearance/traversal causal TE is 3.49/3.45 m (Oxford), 1.98/1.51 (NCLT), and 14.66/7.22 (MCD); pools contain 197/222, 189/214, and 278/232 frames.

Traversal grouping gives lower TE on all three routes; candidate counts differ.

R1.2 [W1/W2, continued]: What does sparse routing add beyond dense temporal search?

Dense temporal search tests the same context without sparse regions: it indexes a 41-frame window at each map position and expands to 220 frames. F³Loc selects Top-4 regions of up to 70 frames. Both use G and the same fine matcher.

Table R2a. Complete causal dense–sparse ablation; mean TE (m).

RouteDense + GF³Loc
Oxford 15-135.184.78
Oxford 17-132.793.45
Oxford 17-143.213.09
NCLT 02-121.561.51
NCLT 02-191.711.83
NCLT 03-311.581.53
NCLT 05-265.775.52
MCD day106.907.22
Macro3.593.62
AuthorsOpenReview ↗

Response to Reviewer JJmF (2/3)

R1.2 [W1/W2, continued]: Matched dense-versus-sparse lookup cost

Table R2b. Matched causal B=1 CPU lookup (ms/query), four threads, same 128 queries/map.

MapDense temporal+GF³Loc
Oxford416.89113.160
NCLT277.6519.581
MCD64.0714.319

Dense/sparse macro TE is 3.59/3.62 m. Sparse routing uses about 40× fewer coarse keys and gives approximately 15–32× faster CPU lookup under the matched B=1 protocol; this is not an end-to-end speedup. Both methods retain the map frame bank.

R1.3 [W2]: What is timed, and are the hardware and batch sizes matched?

We apologize for the earlier GPU-lookup label: both HNSW and F³Loc retrieve on CPU. All updated timings use 128 identical queries/map, a Xeon Gold 6326 with four CPU threads, RTX 4090 encoding, one warm-up and three interleaved passes.

Table R3. Updated matched timing (ms/query); all listed methods use calibrated descriptors.

MapBExhaustive+G lookupHNSW+G lookupF³Loc lookupHNSW+G full-queryF³Loc full-query
Oxford1413.71315.81313.16034.23132.250
Oxford16104.4334.4954.9446.4276.899
NCLT1273.97615.9899.58134.11428.681
NCLT1668.7964.5534.0006.6096.147
MCD161.30713.2024.31931.69923.393
MCD1615.4633.7002.7335.7154.725

Lookup measures CPU search and frame matching; full-query additionally includes encoding/calibration and online transfers. B=1 reports single-query latency and B=16 amortized time/query; offline map and index construction are excluded. On MCD, F³Loc improves route-wide TE (7.22 vs. 12.16 m) and matched B=16 full-query time (4.725 vs. 5.715 ms/query) over HNSW+G. On Oxford, HNSW+G is faster at B=16 (6.427 vs. 6.899 ms/query). Revised Table 13 compares all Table 2 retrieval baselines on matched hardware, with B=1/B=16 CPU lookup and full-query times; its IVF-Flat is the original single-cell control, not the newly matched four-list IVF. The old 476.4 ms measurement included CPU exact search (440.35 ms) and encoder time (34.76 ms). The newly profiled Flat+G B=1 run below is a separate measurement, not a paired speedup.

Table R3a. Oxford 17-13, B=1, CPU Flat+G breakdown (ms/query).

StageTime
BEV to GPU0.233
Query encoder (GPU)19.218
L2 normalization0.043
Spectral calibration G0.410
Descriptor to CPU0.050
CPU exact full-map search412.527
MapPose readout0.034
Other timing/synchronization0.067
Segmented total432.583

CPU exact search dominates Flat+G; the original 476.4 ms used a different protocol. Symmetric retrieval is offline: compute time excludes waiting for 50 future scans, whereas causal uses no future frames. We have added this clarification to the revised Table 1 caption.

R1.4 [W2]: Why do sequence baselines perform poorly, and which models were evaluated?

The SeqLPD-, SeqSLAM-, and SeqNet-style controls test routing on the same frozen BEVPlace++ LiDAR bank, rather than reproducing the full original systems. Their high TE may reflect the mismatches in Fig. 2: sequence aggregation can reinforce coherent distant aliases, while unequal evidence granularity or separate stage mappings can distort coarse rankings. Table R2 compares SeqSLAM-style and F³Loc under the same G, causal/symmetric windows and comparable candidate budgets; F³Loc achieves lower TE on Oxford, NCLT and MCD. The released SeqNet uses image/NetVLAD inputs; our LiDAR S5/S1 heads were source-trained (30 epochs, 300 updates/epoch, batch 128; Oxford/NCLT TE 32.35/12.80 m at 220 comparisons). Official released checkpoints: BEVPlace++ (KITTI 00), EgoNN (MulRan/Apollo), LCDNet (SemanticKITTI), HOTFormerLoc (Oxford); see R1.7 for its Oxford gaps.

Official SeqOT (NCLT-trained, with range images projected from LiDAR scans) gives Oxford 17-13 R@1@5m 66.49%, mean TE 93.04 m, and 7,755 errors >50 m on 42,281 queries. It uses future frames and was not retrained on Oxford. SeqOT is separate from our matched-bank controls.

R1.5 [W3]: How were the settings selected, and how sensitive are the gains?

The encoder and G are fitted on source data. Routing settings were selected using NCLT 02-12 and NCLT 03-31 as development routes, then applied unchanged to Oxford and MCD: causal/symmetric W=40/50, Top-4, and 50-frame cores. Source mixture and LoRA/full-fine-tuning choices were also informed by NCLT 02-12 and NCLT 03-31. We additionally report one-at-a-time sensitivity on NCLT 02-12, plus a separate Sejong validation experiment. The latter independently selects only W and region-core length with Top-K fixed at 4; it does not retroactively describe how the submitted settings were chosen.

AuthorsOpenReview ↗

Response to Reviewer JJmF (3/3)

R1.5 [W3, continued]: Sensitivity and independent Sejong validation

Table R4. Routing sensitivity on NCLT 02-12.

ParameterTested valuesMean TE range (m)
Symmetric radius W10, 20, 30, 40, 501.51–1.63
Top-K2, 4, 6, 81.39–1.68
Region core40, 50, 601.48–1.62

Top-K=2 gives 1.39 m on NCLT 02-12, but Top-K=4 retains more geographic candidates. We apply the same K to every evaluation route rather than choosing route-specific optima.

As an independent post-submission check (not the source of the submitted hyperparameters), we used a separate MulRan pair: Sejong01 as the map and Sejong02 as queries (20,143 eligible queries), distinct from the eight target evaluation routes. With the encoder and G fixed and Top-K=4, we selected W and region-core length by validation mean TE, then evaluated the selected configurations on all eight target routes without further tuning.

Table R4a. Submitted versus validation-selected configurations, with Top-K fixed at 4. Mean TE is macro-averaged equally across eight target routes.

ConfigurationContextWTop-KRegion coreMean TE (m)
Raw flatSingle frame–––6.69
Flat + GSingle frame–––4.53
Submitted F³LocCausal404503.62
Validation-selectedCausal504503.67
Submitted F³LocSymmetric504503.28
Validation-selectedSymmetric504603.41

On Sejong, submitted settings are within 0.43%/1.86% of the best tested C/S Top-4 configurations. Independently selected settings retain 3.67/3.41 m eight-route macro TE (Flat + G: 4.53 m).

R1.6 [W4]: What is new beyond existing components, bilinearity, and ordinary whitening?

Our main contribution is to identify how coarse routing can irreversibly discard the correct geographic hypothesis, and to show how shared comparison geometry and aligned evidence support reduce this risk. F³Loc provides a simple yet effective implementation of this principle. Its contribution lies in the retrieval failure it exposes and addresses, not in averaging, chunking, or PCA individually.

Oxford coarse/fine geometry swaps in manuscript Table 16 give TE 9.36, 12.12, 5.23, and 3.43 m for raw/raw, raw/G, G/raw, and G/G. A better fine metric alone cannot repair a discarded location. Eq. 11 connects regional scores to the calibrated frame comparisons used in final matching; the swaps test how this shared rule affects pruning.

Table R5. Source calibration; eight-route macro TE (m), matched dimensions/normalization.

CalibrationFlatHierarchy CHierarchy S
Sample-weighted pooled PCA whitening4.543.913.55
Shared G4.533.623.28

Unlabeled-map whitening is a competitive alternative. Under symmetric context, it achieves mean TE of 2.36 m on Oxford 17-13 and 1.37 m on NCLT 02-12, compared with 2.15 m and 1.51 m for source-only G.

A fixed source G can be reused across deployment maps without refitting. If map-fitted whitening is re-estimated after map expansion, stored descriptors, region summaries, and search indices must be updated consistently; a fixed source G avoids this calibration-induced reprocessing.

R1.7 [Others]: Clarity, main-text evidence, Table 1 gaps, and ICP time

Presentation. The blue, repeated “Note” at the beginning of Fig. 3’s caption has been removed. We simplified the abstract and introduction, first explaining the practical routing decision before the three mismatches.

Main-text evidence. Fig. 5 now shows the coarse/fine geometry swaps in the main text; the full candidate-coverage and fixed-pool ranking results remain in the appendix. Fig. 6 includes the MCD long-tail results. We added “Shared geometry and temporal routing” (Table 4), separating G, dense temporal evidence, and sparse routing, and “Maps beyond a single traversal” (Table 5), covering multiple NCLT recordings and map-pose regions. These central results are now visible in the main paper. Appendix E’s sensitivity sweeps are too extensive to place in full in the main paper; instead we added the independent Sejong check in Appendix E.1.

Table 1 gaps. HOTFormerLoc uses the official Oxford-trained checkpoint. Oxford evaluation on this checkpoint would be in-domain, not held-out transfer; we report only its NCLT transfer results (TE 33.51/45.64/32.39 m; 24.6 ms). We did not run Oxford accuracy or timing, so Table 1’s “–” and “–/24.6” denote unreported Oxford measurements, not zero scores.

GeoRefine. Its Oxford/NCLT 224.4/284.7 ms totals include separately timed Fast retrieval and RANSAC/ICP refinement. F³Loc-Fast returns MapPose without refinement.

Please let us know if any questions or concerns remain.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall Author Response — F³Loc

We sincerely thank the Area Chair and all reviewers for their careful reading, constructive feedback, and thoughtful assessment of our work. We have revised the manuscript and provided detailed, experiment-backed responses to the concerns raised.

  • To JJmF: We separated the contributions of spectral calibration and hierarchical retrieval across eight routes, added matched-temporal and dense-versus-sparse controls, and remeasured retrieval latency under matched hardware. We also clarified our novelty in addressing irreversible geographic hypothesis pruning through aligned retrieval geometry and granularity. We further clarified the baseline checkpoints and missing HOTFormerLoc results, reported PCA and map-fitted whitening comparisons, documented parameter selection with an independent Sejong check, and brought key evidence into the main text.

  • To gnME: We expanded the fixed-calibration comparisons, including sample-weighted PCA and a second descriptor encoder, and added matched-context baselines and timing analyses. We further reported median and tail errors, harmful and beneficial ordering inversions under the original XY protocol, and tests involving multi-session maps, map gaps, and different observed speeds.

  • To RUqg: We evaluated multi-traversal maps using temporal and map-pose region grouping, tested shuffled submaps and missing map sections, and clarified the properties required of a spatial partition. We also compared F³Loc with HNSW+G using matched CPU hardware, query sets, batch sizes, and timing scopes.

We appreciate the opportunity to strengthen the paper and welcome any further questions or concerns. We would be happy to clarify any remaining points during the discussion.

Area_Chair_tV3VOpenReview ↗

reminder

Dear reviewers,

This is a gentle reminder to read the authors’ rebuttals and respond in the discussion. The authors have put considerable effort into addressing your feedback, and we encourage you to acknowledge their responses and follow up on any remaining concerns.

Thank you for your time and effort throughout the review process, and for your contributions to the community.

Best Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.