Natural-language point-cloud localization typically retrieves candidate map cells and then predicts a position inside each candidate, yet the final output is still determined by retrieval order. A retrieval score measures region-description relevance, not whether the coordinates a region yields are accurate: a relevant region may produce a displaced prediction, and a lower-ranked region may contain a successful one. We study the final selection problem after the candidate pool and Fine predictions are fixed---how to choose the coordinate most likely to localize successfully, and under which conditions the gain holds. We propose SCAR (Spatial Consistency-Aware Reranking). SCAR treats each cell--Fine-coordinate pair as a localization hypothesis, interprets directional relations as functions of position, and reads spatial support at the backbone's actual Fine point (Fine-conditioned Evidence Readout). It organizes this readout with center support, semantic presence, retrieval score, and geometric offsets into a 23-D representation, trains a shared scorer with localization-success labels, and at inference reranks existing hypotheses and returns their original coordinates. Experiments cover four public backbones and a corrected Des4Pos reimplementation across five domains. Excluding one protocol-reference configuration, exploratory TEST point estimates on the remaining 24 configurations show Top-1@5,m above Original on 23 and Cell R@1 above Original on all 24. On development-set VAL, the Fine readout improves Top-1@5,m by 5.77 percentage points (pp) over a center surrogate on corrected Des4Pos*KITTI, and the complete scoring scheme improves it by 13.98,pp over FSR on MambaPlace×nuScenesFull; the latter measures the complete scheme, not learned ranking alone.
Figures & tables
Fig. 1: Candidate availability and selection on MNCL × KITTI VAL with a fixed Top-25 prediction pool. (a) Original and Oracle@25 Top-1@5 m. (b) Queries corrected and broken by two training-free rankings relative to Original. The oracle uses GT only to measure recoverable coverage. Query counts, net gains, and the paired interval are discussed in the text.
Fig. 2: SCAR overview. A frozen Coarse–Fine backbone supplies 25 cell–position hypotheses. Query relations are matched to map objects, spatial support is read at each existing Fine prediction, and the resulting 23-D representation is normalized with TRAIN statistics and scored by a shared MLP. SCAR reranks the fixed hypotheses and returns the existing Fine coordinate of the selected candidate. Point-cloud layouts, relations, support fields, and scores are illustrative.
Benchmark
TRAIN
VAL
TEST
Split-unit definition
KITTI360Pose
28,807
3,187
11,505
30 m cells; six descriptions; TEST covers three scenes
PandaSetPose
4,000
2,000
2,000
six descriptions per pose; nearest retained anchor cell
nuScenesPose Full
15,113
7,635
7,403
2–6 original descriptions; 210 TEST scenes with queries
S3DISPose
1,084
544
542
rooms as split unit; four descriptions; TEST 68 rooms
ParisLillePose/NPM3D
120
128
116
20 m cells; three descriptions; TEST 29 query scenes
TABLE I: Frozen localization splits. Counts are localization requests; one request may contain several descriptions.
KITTI
S3DIS
PandaSet
Backbone
1 m
3 m
5 m
10 m
1 m
3 m
5 m
10 m
1 m
3 m
5 m
10 m
Text2Loc
6.46 (+3.36)
37.16 (+16.68)
59.45 (+24.31)
75.77 (+26.88)
30.26 (+20.85)
39.30 (+27.31)
41.33 (+27.49)
41.70 (+27.31)
7.70 (+0.15)
30.85 (+2.30)
44.95 (+4.00)
55.25 (+4.75)
Text2Loc++
8.29 (+3.78)
41.23 (+15.09)
64.15 (+20.43)
78.86 (+21.10)
34.50 (+18.27)
44.10 (+24.17)
45.94 (+25.28)
47.97 (+26.20)
4.55 (+0.35)
29.45 (+4.00)
44.30 (+6.95)
56.30 (+8.85)
MNCL
7.20 (+2.83)
40.40 (+11.26)
63.82 (+14.23)
79.69 (+13.99)
22.88 (+18.08)
36.90 (+29.70)
38.38 (+30.44)
39.11 (+31.18)
4.65 (-0.30)
27.20 (-2.10)
42.05 (-0.15)
52.50 (-0.10)
MambaPlace †
7.21 (+3.95)
36.01 (+17.49)
57.07 (+25.09)
72.92 (+28.34)
22.14 (+16.24)
38.19 (+27.68)
40.77 (+28.78)
42.07 (+28.60)
1.80 (+0.45)
14.35 (+5.65)
22.15 (+8.25)
34.65 (+11.85)
Des4Pos (reimpl.)
6.62 (+4.17)
36.11 (+19.06)
57.84 (+28.47)
73.53 (+32.37)
22.32 (+14.21)
36.35 (+22.51)
37.82 (+23.25)
39.11 (+23.25)
3.25 (+0.60)
21.60 (+8.95)
31.55 (+11.20)
44.65 (+14.65)
TABLE II: TEST localization success (%) of SCAR at four distance thresholds. Each cell shows the success rate above and the difference from Original (pp) below. The primary threshold is 5 m.
KITTI
S3DIS
PandaSet
Backbone
R@1
R@3
R@5
R@10
R@1
R@3
R@5
R@10
R@1
R@3
R@5
R@10
Text2Loc
43.74 (+13.90)
70.00 (+19.85)
76.99 (+16.82)
81.33 (+10.01)
41.70 (+27.31)
63.84 (+32.10)
72.51 (+25.83)
85.42 (+20.66)
52.35 (+7.30)
77.40 (+5.20)
85.40 (+2.40)
92.90 (+2.80)
Text2Loc++
45.02 (+10.02)
72.36 (+14.74)
80.15 (+13.26)
85.02 (+8.09)
48.15 (+26.20)
70.11 (+24.54)
82.29 (+24.91)
89.11 (+18.08)
49.10 (+7.95)
69.15 (+4.45)
83.20 (+9.00)
91.75 (+2.65)
MNCL
47.41 (+3.51)
76.62 (+9.77)
84.19 (+9.38)
88.80 (+5.77)
39.30 (+31.18)
60.15 (+38.38)
68.82 (+35.06)
77.86 (+25.46)
47.85 (+1.80)
74.20 (+5.95)
82.45 (+5.10)
89.55 (+2.50)
MambaPlace †
40.83 (+16.28)
65.34 (+21.56)
72.03 (+18.83)
76.26 (+11.79)
42.25 (+28.78)
64.76 (+33.03)
73.80 (+31.55)
82.66 (+23.99)
29.50 (+9.70)
46.90 (+11.95)
57.35 (+11.55)
73.20 (+14.85)
Des4Pos (reimpl.)
43.11 (+19.44)
65.72 (+23.06)
71.67 (+20.11)
75.35 (+12.07)
39.11 (+23.06)
61.44 (+27.31)
72.51 (+29.15)
82.10 (+18.82)
41.55 (+13.45)
61.70 (+8.10)
72.05 (+7.85)
83.25 (+6.35)
TABLE III: TEST Cell R@1, R@3, R@5, and R@10 (%) of SCAR. Each cell shows the recall above and the difference from Original (pp) below.
Dataset
Backbone
Original
Semantic Matching
Rank Fusion
FSR
SCAR
KITTI
Text2Loc
35.14 / 29.84
49.05 / 41.28
54.58 / 44.62
58.65 / 46.51
59.45 / 43.74
Text2Loc++
43.72 / 35.00
56.45 / 44.98
59.05 / 46.05
63.36 / 47.51
64.15 / 45.02
MNCL
49.59 / 43.90
54.90 / 46.01
57.56 / 46.62
62.00 / 48.19
63.82 / 47.41
MambaPlace
31.98 / 24.55
45.31 / 35.74
53.72 / 42.36
57.65 / 44.18
57.07 / 40.83
Des4Pos (reimpl.)
29.37 / 23.67
46.03 / 36.71
53.01 / 42.67
57.41 / 44.37
57.84 / 43.11
S3DIS
Text2Loc
13.84 / 14.39
28.23 / 29.89
36.90 / 38.38
44.83 / 46.49
41.33 / 41.70
TABLE IV: Same-pool TEST comparison of Semantic Matching, Rank Fusion, FSR, and SCAR. Each entry reports Top-1@5 m / Cell R@1 (%).
Fig. 3: TEST Top-1 localization success as a function of the true-distance threshold (1, 3, 5, and 10 m). All five panels use the Text2Loc backbone and its fixed Top-25 candidate pool and Fine coordinates on KITTI, PandaSet, nuScenes, S3DIS, and NPM3D. Original, FSR, and SCAR are three rankings of the same candidate set. The vertical axis is 0–80% in every panel; the gray dashed line marks the 5 m primary metric. Curves connect observed values without smoothing. All TEST comparisons are exploratory; Table II reports the four-threshold SCAR results, and Supplement Table S4 gives additional comparisons.
Fig. 4: Real corrected Des4Pos × KITTI VAL cases, shown from left to right: (a) correction, (b) preservation, and (c) degradation. Stars denote GT, orange triangles Original predictions, and purple squares SCAR predictions; dashed circles have a 5 m radius. Coordinates are relative to GT, with the same scale in all panels. Errors below each panel are Original → SCAR. The two predictions coincide in (b). Case identities, query descriptions, and selection details are given in Supplement S4.3.
Fig. 5: VAL Top-1@5 m differences between F and each control: (a) corrected Des4Pos × KITTI and (b) MambaPlace × nuScenesFull. Bars show point estimates; whiskers are paired 95% cluster-bootstrap intervals. Filled bars compare trained F/C/N arms, while hatched bars compare the complete F scoring scheme with training-free FSR. Both panels use scorer seed 0 and the same horizontal scale. Complete results are in Supplement Tables S8 and S9.
Fig. 6: Spatial-support distributions on MNCL × KITTI VAL. (a) Effective matched-relation count m , with the dashed line marking the m≥4 gate. (b) Joint histogram of center and Fine support among active-gate pairs; color indicates pair count on a logarithmic scale, white bins are empty, and the diagonal marks equal support. (c) Changes in satisfied-relation count at the Fine point relative to the center, normalized separately within 5 m-hit and 5 m-miss candidate pairs. These are descriptive distributions, not ranking ablations.
Method
Native K10
Same-pool K25
FSR
SCAR
Text2Loc
126.40
275.99
279.68
276.39
Text2Loc++
93.69
207.64
212.77
214.89
MNCL
252.87
584.45
588.88
592.78
MambaPlace
Not measured
Des4Pos (historical impl.)
100.49
191.98
199.65
196.25
TABLE V: KITTI fixed-VAL-query median stage-wise computational latency (ms) under warm cache. Text2Loc native K10/fixed K25 is 126.40/275.99; MambaPlace was not measured. The Des4Pos row is the historical implementation; its corrected reimplementation was not measured.
Configuration
Coarse
Fine
Evidence type
MambaPlace × KITTI
8 h 37 min 09 s
14 h 03 min 26 s
Start/end wall clock
MambaPlace × PandaSet
1 h 08 min 26 s
1 h 45 min 50 s
Session transcript; log not archived
MambaPlace × nuScenes
5 h 30 min 01 s
8 h 18 min 01 s
Start/end wall clock
MambaPlace × S3DIS
27 min 22 s
31 min 11 s
Start/end wall clock
MambaPlace × NPM3D
2 min 06 s
3 min 52 s
Start/end wall clock
Text2Loc++ × S3DIS
47 min 44 s
45 min 23 s
Start/end wall clock
TABLE VI: Stage training durations in the existing logs. Durations correspond to the original run environment and recipe of each configuration and exclude failed attempts and overall project development overhead.
Language-based 3D localization retrieves the point-cloud submap containing a target position from descriptions of nearby objects and their spatial relations. Existing methods typically compress queries and submaps into global descriptors, potentially obscuring object-level semantics and cross-description spatial coherence. We propose Position-Conditioned Evidence Localization (PosEviLoc), a query-position-aware framework for coarse text-to-point-cloud localization. Instead of relying on global matching, PosEviLoc evaluates each candidate submap using explicit semantic and spatial evidence. It models direction as a relation jointly determined by an object position and a hypothetical query position. The resulting Query-Position Spatial Evidence Field (QSEF) measures the fraction of query descriptions supported at each hypothetical position, explicitly capturing their agreement without using the ground-truth query pose to construct the evidence field. A Multi-Level Evidence Readout (MER) summarizes this evidence in a compact representation, which a lightweight MLP converts into a retrieval score. Across five benchmarks, PosEviLoc outperforms MNCL by an average of 17 percentage points in Recall@1. When used as a plug-and-play reranker, it improves MNCL by an average of 16 percentage points. Moreover, PosEviLoc introduces substantially fewer parameters and achieves faster inference speed than existing methods.
Tianyi Shang, Yike Shi, Zhenyu Li
Purdue University · Xi’an Jiaotong-Liverpool University · Qilu University of Technology (Shandong Academy of Sciences)
Text-to-point-cloud localization estimates a position in a city-scale 3D map from descriptions of surrounding objects. Existing coarse-to-fine methods retrieve submaps using aggregate learned compatibility and then localize within a selected submap. However, repetitive or similar urban objects can inflate the embedding similarity between the query and multiple submaps, even when the instance layout within a submap violates the query description. Meanwhile, query-relevant instances often span submap boundaries, leaving the retrieved submap with incomplete contextual evidence. We term these failure modes layout-inconsistent aliasing and boundary evidence incompleteness, respectively. To address them, we propose PARC-Loc, a coarse-to-fine localization framework built on Partial Assignment with Relational Consistency (PARC). PARC jointly models hint-object compatibility and pairwise spatial relations, allowing unmatched elements while favoring assignments consistent with the queried layout. At the coarse stage, its candidate-level assessment complements neural similarity for layout-consistent submap selection. At the fine stage, the context is expanded with query-relevant instances from adjacent submaps, while PARC yields object-level matching weights that guide cross-modal attention. Extensive experiments on KITTI360Pose and CityLoc show that PARC-Loc outperforms conventional coarse-to-fine baselines. On KITTI360Pose, our method improves Top-1 localization recall at 5 m from 0.50 to 0.67, achieving a 34% relative gain over the strongest baseline.
Scene Coordinate Regression (SCR) has recently emerged as a promising approach for LiDAR-based localization, achieving accurate localization without requiring an explicit 3D map. Despite their effectiveness, existing SCR methods rely on scene classification-based global embedding that struggles to provide fine-grained discrimination among nearby locations. Moreover, their reliance on uniform sampling of local features during training assigns equal importance to all points, thereby inadvertently propagating features from dynamic objects or unstable regions and potentially degrading training stability. In this paper, we present ReLoc, a revamped SCR architecture that can effectively address these limitations. First, we redesign the global embedding module by combining learnable context tokens with a feature aggregator to capture richer and more discriminative scene context. Second, we introduce an attention-based local feature enhancement module to mitigate the impact of noisy local features while encouraging context-consistent structures, yielding more robust local feature representations. Experimental results on two large-scale outdoor datasets demonstrate that our approach achieves state-of-the-art accuracy over previous SCR-based methods while maintaining real-time inference performance.
Heejoon Moon, Yurim Cho, Je Hyeong Hong
Department of Artificial Intelligence, Hanyang University · Department of Electronic Engineering, Hanyang University