We propose PadLoc, a training-free localization back-end that places a moving sensor within a known floor plan. Floor plans are cheap, persistent, and robust to appearance change, but recent learned-depth methods localize against them by filtering per-frame depth through dense (x,y,θ) pose grids. PadLoc keeps the swappable depth front-end and replaces the dense grid back-end with sparse geometric matching: dominant wall orientations define a small pose-candidate set, masked normalized cross-correlation scores observed free space against the plan, and a moving-pose Gaussian-mixture Bayes filter plus consensus readout decodes the trajectory. Holding each competitor's depth fixed, PadLoc matches or beats UnLoc's forward filter and the F3Loc filter at every Gibson trajectory length, and UnLoc's offline refinement from T=35 on; it exceeds Z-FLoc's published results with a shared MapAnything front-end. On Bonn LiLocBench, PadLoc improves pooled SR@1m over UnLoc by 23-32 pp under wheel odometry and 35-43 pp with ground-truth motion; with full-FOV LiDAR plus wheel odometry the same readout reaches 100% SR@1m at 9-10 cm median error on the static sequence. On the Hilti-Trimble Localization Challenge PadLoc localizes a camera-only fisheye rig on construction floors hundreds of metres across at 0.90-0.98 online SR@2m on three of five sessions; adding learned wall isolation, applied to every session, covers all five at 0.67-0.92 with the localizer unchanged throughout. PadLoc thus matches or outperforms dense grid back-ends while preserving sensor modularity and a compact state that does not grow with trajectory length.
Figures & tables
Figure 1: PadLoc method overview on real data (tiles: intermediates of one trajectory chunk). A known floor plan, depth profiles, and planar odometry enter (1); the training-free back-end fuses a short window into a target and observation mask, which bounds drift (2), reduces rotation to four wall-angle heading candidates (3), scores translations with masked NCC and keeps top- K peaks (4), propagates K≤10 gated Gaussian hypotheses with odometry (5), and re-reads the raw NCC fields by consensus to choose the global anchor and recover poses (6; dashed: ground truth).
Figure 2: Sliding-window matching and consensus decoding ( cons ). Columns show one trajectory from early ambiguity to the final anchor. Rows show the accumulated target, masked-NCC top- K votes, GM-BF hypotheses, and the cumulative consensus readout.
T=15
T=20
T=35
T=50
T=100
SR ↑
R ↓
SR ↑
R ↓
SR ↑
R ↓
SR ↑
R ↓
SR ↑
R ↓
F3Loc mono depth
F3Loc filter
0.291
0.27
0.345
0.23
0.550
0.21
0.705
0.18
0.892
0.18
PadLoc frozen
0.496
0.29
0.585
0.28
0.766
0.27
0.821
0.24
0.946
0.28
F3Loc comp_s depth
F3Loc filter
0.366
0.21
0.410
0.18
0.622
0.16
0.795
0.14
0.946
0.15
Table 1: Matched-depth back-end comparison on Gibson: last-10 SR@1 m ( ↑ ) and, following UnLoc’s protocol, the mean last-10-frame RMSE over each method’s own successful trials (m, ↓ ; not a precision ranking, Tab. 2 ), pooled over 9 scenes. Each F3Loc/UnLoc block fixes the depth input; F3Loc filter and UnLoc fwd use the per-frame, PadLoc the frozen readout (§ 4.1 ; chunk-end: supplement). Bold: best SR per comparison (frozen vs filter, accum vs batched or refined rows), ties bold. Z-FLoc quoted from its paper (‘–’).
T=15
T=20
T=35
T=50
T=100
UnLoc fwd
.573/.17
.610/.14
.771/.11
.798/.11
.822/.12
UnLoc refined
.670 /.18
.717 /.17
.765/.16
.791/.14
.783/.17
PadLoc frozen
.601/.20
.695/.17
.774/.15
.839/.15
.811/.19
PadLoc accum
.667/.18
.715/.18
.809 /.14
.856 /.13
.846/.14
PadLoc frozen, 5cm
.624/.18
.694/.15
.809 /.12
.849/.11
.887 /.11
Table 2: Unconditional metrics on the Gibson UnLoc- μ block of Tab. 1 : SWA ↑ / all-trial median last-10 RMSE (m) ↓ ; bold = best SWA per column (Bonn: supplement).
T=15
T=35
T=100
SR
R
SR
R
SR
R
back-end build-up
per-frame NCC argmax
0.108
3.11
0.081
3.21
0.216
2.37
+ accumulation (argmax)
0.817
0.17
0.928
0.14
1.000
0.13
+ consensus, gate/heal off
0.687
0.21
0.856
0.15
0.919
0.14
+ GM-BF gate & heal
0.731
0.20
0.919
0.15
1.000
0.19
Table 3: Component ablation (last-10 SR@1 m ↑ / median last-10 RMSE (m) ↓ , per-chunk reset; consensus rows frozen). Top: Gibson build-up and leave-one-out (UnLoc μ depth, n=268/111/37 ; full config bold). Bottom: the build-up under real drift (Bonn static_0 , D455 + VO, n=59/29/19 ). Accumulation-alone is the offline aggregate-then-match reference: the frozen decoder is within one chunk of it at T≥35 , while under drift the gate keeps the readout near the reference (row 3 vs 4) and above the smeared batch target.
UnLoc
PadLoc cons , wheel
scene
150
300
450
150
300
450
static_0
0.32
0.59
0.68
0.59
0.79
0.89
dynamics_0
0.33
0.43
0.60
0.67
0.71
0.60
lt_changes_0
0.21
0.43
0.21
0.60
0.71
0.79
lt_ch_dyn_0
0.33
0.44
0.39
0.45
0.63
0.72
pooled SR@1 m
0.30
0.49
0.46
0.56
0.71
0.79
Table 4: PadLoc vs UnLoc on Bonn (monocular structural depth and camera-built structural map held fixed; UnLoc with its own b^ ). Last-10 SR@1 m at T=150/300/450 per sequence and pooled; PadLoc frozen, UnLoc per-frame readout (chunk-end: supplement); PadLoc under wheel odometry, UnLoc under its ground-truth transitions. PadLoc wins 11 of 12 cells (tie: dynamics_0 , T=450 , n=5 ). Pooled n=172/84/56 . 16 rays: PadLoc fed only the 16 DeSDF rays UnLoc’s likelihood reads from the same profile.
session
T
SR@2 m
mean
score
30/30
az55
+ seg
(m)
floor_1
3923
0.977
0.710
0.718
0.90
68.1
run_2
2569
0.907
0.953
0.895
0.68
74.3
05-05
5594
0.898
0.813
0.920
1.11
62.3
run_1
3067
0.456
0.689
0.689
2.68
32.3
UG1
6124
0.156
0.000
0.668
5.62
0 9.9
Table 5: Hilti large-building localization (camera-only fisheye, MapAnything depth, OKVIS2-X prior). Online SR@2 m ( ↑ ) for the batch-matched 30/30 window and front-end arms (az55: ±55∘ side crops; + seg: learned wall isolation), back-end unchanged; each column is one configuration applied to all sessions, none is selected per session; mean error/score at 30/30 . run_2 , 05-05 , run_1 : the three floor_2 sessions (2025-10-28 run 2, 2025-05-05, 2025-10-28 run 1).
Figure 3: Matcher cost vs. floor-plan size ( 10cm grid; RTX 3090). One UnLoc call scores all 36 orientations per frame; a PadLoc window scores all B≈4 headings (solid; one heading dashed); complete back-end cost is given in the text. UnLoc grows 2.3→209\text{,}\mathrm{ms} ; a PadLoc window grows $3.5\!\to\!$13.7\text{\,}\mathrm{ms} ( 15× at mall scale, 76× per frame at stride 5 ; UnLoc is cheaper at home scale).
Appendix figures & tables25 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : Observation semantics on the blurred reference. (a) σ -blurred reference field. (b) A window target overlaid: yellow wall hits register on the blue reference bands, teal cells are observed free space, and unobserved cells drop out of the masked NCC.
Figure S2 : Orientation pre-pass mechanism. Hough wall segments from the reference and accumulated target scan are converted into length-weighted wall-angle histograms. Cross-correlation aligns the target histogram to the reference axes and yields the directed heading set plus 180∘ mirrors (here 4.0∘/94.2∘ ), which masked NCC scores downstream.
Figure S3 : Cumulative top- K NMS recall of the GT pose on the headline consensus decoder’s surface (moving-pose GM-BF + consensus-sum, chunk-end readout, GT depth, 10cm grid, 37 gibson_t chunks; GT-hit and NMS separation 0.30m , quantization-fair at 10cm ). σ is per-frame odometry noise: a seeded random walk of σ cm on each translation axis and σ∘ on heading per frame. At σ=2.0 the global argmax recovers GT in only 62.2% of chunks, while the deployed K=10 (dashed) recovers 83.8% (top- 20 : 86.5% ); for σ≤0.5 the argmax misses at most one chunk (at σ=0 one chunk is missed at every K ).
parameter
value
set by
fixed across all datasets
wall-hit mask weight wwall
55
Gibson
b^ column-drop threshold
0.30m
Gibson
overlap-rejection threshold
0.7
Gibson
heading-overlap exponent
0.70
own sweep (unique opt.)
top- K peaks / mixture cap
10
Gibson
Appendix
Table S1 : Principal hyperparameters. Top: fixed across all datasets. Bottom: front-end-determined, set by stated rules.
setting
T=15
T=20
T=35
T=50
T=100
published ( w=55 , τ=0.30 )
.731
.840
.919
1.000
1.000
wwall=25
.746
.835
.910
.974
1.000
wwall=40
.728
.840
.928
1.000
1.000
wwall=70
.713
.835
.937
1.000
1.000
wwall=100
.709
.815
.928
.987
1.000
b^ cutoff τ=0.20
.705
.815
.937
.974
1.000
Appendix
Table S2 : Sensitivity sweeps (UnLoc μ depth, 9 scenes, pooled last-10 SR@1 m, frozen readout). One constant is varied at a time from the published setting (bold); last row: the pair the test-set sweeps favoured with the re-summed consensus (Sec. P ).
T=15
T=35
T=100
W/S
SR
R
SR
R
SR
R
5/1
0.683
0.24
0.901
0.16
0.973
0.21
5/5 (deployed)
0.731
0.20
0.919
0.15
1.000
0.19
10/5
(0.769)
0.20
0.910
0.15
1.000
0.17
10/10
(0.761)
0.21
0.892
0.16
1.000
0.13
15/5
(0.817)
0.17
0.928
0.16
1.000
0.19
Appendix
Table S3 : Fusion window/stride sweep on Gibson (UnLoc μ depth, 9 scenes, last-10 SR@1 m ↑ / median last-10 RMSE (m) ↓ , frozen readout; n=268/111/37 ; at T=15 , windows longer than five frames emit their first pose after scoring starts and earlier scored frames take that pose ( 15/5 : one window, frozen = chunk-end), in parentheses). The operating point sits on a plateau at T≥35 .
T=150
T=300
T=450
SR
R
SR
R
SR
R
LiDAR 270∘ + wheel (clean)
per-frame NCC argmax
0.542
0.20
0.724
0.12
0.526
0.21
accumulation → full
1.000
0.08
1.000
0.09
1.000
0.10
D455 RGB-D + VO (high-drift)
per-frame NCC argmax
0.153
7.7
0.069
7.7
0.105
10.3
Appendix
Table S4 : Component ablation on Bonn ( static_0 , last-10 SR@1 m ↑ / median RMSE (m) ↓ ; n=59/29/19 chunks at T=150/300/450 ; consensus rows frozen readout, argmax and accumulation rows their single readout). The clean LiDAR + wheel arm saturates (accumulation alone reaches 1.000 ); the high-drift D455 + VO arm separates the components.
Figure S4 : Readout timing on Gibson ( T=15 , W=S=5 , 1-based frames). Bars: windows; dot: pose stamp (end −2 ); triangle: output available after the window’s last frame; dashed line: first output. Grey band: the ten scored frames. Forward filters (F3Loc, UnLoc- fwd ) emit one pose per frame. PadLoc frozen: each scored frame is coloured by the window that serves it, the latest window ended at or before that frame; numbers give the anchor age in frames. PadLoc delayed: all ten frames use the chunk-end anchor read at frame T .
T=150
T=300
T=450
L
SR
p90
SR
p90
SR
p90
0
0.971
0.18
0.883
0.29
0.905
0.76
2
0.974
0.18
0.884
0.28
0.911
0.73
5
0.979
0.18
0.887
0.27
0.918
0.71
10
0.982
0.18
0.898
0.26
0.920
0.68
∞
0.983
0.18
0.916
0.25
0.913
0.53
Appendix
Table S5 : Bounded-lag online readout on Bonn ( static_0 drifty-VO arm: LiDAR 106∘ depth + VO). Finite L uses delayed window outputs; L=∞ is an offline reference. Online SR@1 m over emitted windows / steady-state p90 error (m).
T=15
T=20
T=35
T=50
T=100
SR ↑
R ↓
SR ↑
R ↓
SR ↑
R ↓
SR ↑
R ↓
SR ↑
R ↓
F3Loc mono depth
F3Loc filter
0.511
0.25
0.530
0.24
0.703
0.19
0.808
0.18
0.892
0.18
PadLoc delayed
0.608
0.28
0.695
0.28
0.775
0.26
0.846
0.26
0.946
0.26
F3Loc comp_s depth
F3Loc filter
0.567
0.21
0.580
0.19
0.757
0.15
0.872
0.14
0.946
0.14
Appendix
Table S6 : Main-paper Table 1 under the chunk-end readout (Gibson, last-10 SR@1 m ↑ / mean last-10 RMSE over successes (m) ↓ ). Bold: best SR per comparison (delayed vs filter rows, accum vs batched or refined rows), ties bold.
T=15
T=20
T=35
T=50
T=100
UnLoc fwd
.703/.13
.748/.13
.809/.11
.821/.12
.809/.14
UnLoc refined
.670/.18
.717/.17
.765/.16
.791/.14
.783/.17
PadLoc delayed
.695/.16
.757 /.16
.799/.13
.835/.15
.823/.18
PadLoc accum
.667/.18
.715/.18
.809/.14
.856 /.13
.846/.14
PadLoc delayed, 5cm
.716 /.16
.737/.15
.828 /.10
.854/.11
.887 /.12
Appendix
Table S7 : Main-paper Table 2 under the chunk-end readout (Gibson UnLoc- μ block: SWA ↑ / all-trial median last-10 RMSE (m) ↓ ; bold = best SWA per column).
Gibson ( T=15/35/100 )
+ consensus, gate/heal off
0.806
0.18
0.883
0.14
0.973
0.13
+ GM-BF gate & heal
0.843
0.16
0.937
0.13
1.000
0.18
−θ -overlap weight
0.817
0.18
0.928
0.14
1.000
0.18
−b^ column drop
0.810
0.19
0.910
0.13
1.000
0.17
− consensus heal
0.832
0.18
0.910
0.14
0.973
0.18
Bonn LiDAR 270∘ + wheel ( T=150/300/450 )
Appendix
Table S8 : Consensus rows of main-paper Table 3 and Table S4 under the chunk-end readout (last-10 SR@1 m ↑ / median last-10 RMSE (m) ↓ ; Gibson UnLoc μn=268/111/37 at T=15/35/100 , Bonn static_0 n=59/29/19 at T=150/300/450 ; full config bold).
UnLoc
PadLoc cons , wheel
scene
150
300
450
150
300
450
static_0
0.37
0.59
0.74
0.59
0.79
0.89
dynamics_0
0.33
0.43
0.60
0.67
0.71
0.60
lt_changes_0
0.28
0.43
0.21
0.60
0.71
0.79
lt_ch_dyn_0
0.35
0.41
0.39
0.47
0.63
0.72
pooled SR@1 m
0.34
0.48
0.48
0.56
0.71
0.79
Appendix
Table S9 : Main-paper Table 4 under the chunk-end readout (Bonn, monocular structural depth, cameras-only structural map; last-10 SR@1 m; PadLoc under wheel odometry, UnLoc- fwd under its ground-truth transitions). PadLoc wins 11 of 12 cells (tie: dynamics_0 , T=450 , n=5 ). 16 rays: equal-ray control as in main-paper Table 4. Replacing UnLoc’s ground-truth transitions by wheel odometry leaves its per-frame SR@1 m unchanged in 20 of 22 cells and one chunk higher in the other two (constant- σ head, shipped and cameras-only maps).
T=300
T=450
depth
motion
SR
R
SR
R
LiDAR 270∘
VO
1.000
0.10
0.947
0.12
LiDAR 270∘
wheel
1.000
0.09
1.000
0.10
LiDAR 106∘
VO
0.931
0.10
0.947
0.12
D455 RGB-D
VO
0.966
0.22
0.842
0.31
LiDAR 270∘ , σ=0.5
VO
0.966
0.09
0.895
0.14
Appendix
Table S10 : Bonn static_0 modularity (last-10 SR@1 m ↑ / median RMSE (m) ↓ , frozen readout, per-chunk reset, n=29/19 ). Top: deployed σ ( 1.0 LiDAR, 0.5 RGB-D; bold = best SR); bottom: the same rows with σ swapped.
Figure S5 : Hilti online pose error ( W/S=30/30 , with-windows reference). Per-frame error ∥x^−xgt∥ vs. trajectory fraction for the five GT sessions; the legend lists frame count and mean error. Blue: localizing sessions; red: run_1 and UG1 ; dashed: 2m threshold.
Figure S6 : Whole-run trajectory on floor_1 . The online per-window estimate (purple) vs. ground truth (green) on the occupancy floor plan; the estimate tracks the full building loop (SR@ 2m0.98 , median 0.81m ).
Figure S7 : Per-window locking on floor_1 (qualitative). Nine of the 24 windows of the first 720 frames. Each panel shows the masked-NCC field (inferno), the depth target at the ground-truth pose (purple), the raw NCC peak (gold star), the decoded estimate (cyan ring), and GT (green plus).
Figure S8 : Hilti failure taxonomy — three regimes. One row per regime. Columns: the two rectified camera views; the wall extraction placed on the plan at the ground-truth (red) and emitted (blue) pose, for a success and a failure window; and the per-window online error ( ±55∘ multi-crop variant; dashed: 2m success threshold). floor_1 fails by a parallel offset of a correct alignment; run_1 ’s off-plan walls localize on free space and fail by aliasing; UG1 reconstructs blobs, and its timeline also shows the recovery with learned wall segmentation ( + seg, blue).
SR@1 m ↑ at T
cell
15
20
35
50
100
10cm
0.731
0.840
0.919
1.000
1.000
5cm
0.743
0.815
0.928
0.974
1.000
3cm
0.679
0.795
0.919
0.962
1.000
2cm
0.638
0.750
0.892
0.949
1.000
1cm
0.511
0.640
0.802
0.910
0.973
Appendix
Table S11 : Sparse depth has a resolution sweet spot (Gibson, UnLoc μ depth, 9 scenes, single frozen config, per-chunk reset). Last-10 SR@1 m ( ↑ ) vs. matching cell; per-column best in bold (ties unmarked). Short- T success peaks at 5cm and falls below it, while T=20 and T=50 favour the coarse grid.
T=300
T=450
cell
SR ↑
RMSE ↓
SR ↑
RMSE ↓
5cm
1.000
0.086
1.000
0.102
10cm
1.000
0.103
1.000
0.095
15cm
1.000
0.121
1.000
0.124
20cm
1.000
0.204
0.947
0.227
Appendix
Table S12 : Dense depth keeps its success rate on coarse cells (Bonn, 270∘ LiDAR + wheel odometry, static_0 ). Last-10 SR@1 m ( ↑ ) / median last-10 RMSE (m, ↓ ) vs. matching cell, frozen readout. SR stays at 1.000 up to 15cm and error grows beyond 10cm ; the native 5cm map floor reproduces the modular lidar+wheel headline; per-column best in bold (ties unmarked).
Figure S9 : Per-call masked-NCC latency (RTX 3090, cuFFT) vs. target-window side on the 30×24 m floor. Solid lines reuse the precomputed reference FFT; dashed lines rebuild it each call.
Figure S10 : On-device power-of-two staircase (iPhone 15 Pro & 17 Pro, vDSP, 10cm , minimum over repeated calls). (a) Fixed 11m target, growing floor plan; (b) fixed home plan, growing target. Both phones step through the same power-of-two FFT sizes and are flat within each size; the 17 Pro is ∼1.2× faster.
Gibson, F3Loc depth
T=15
T=20
T=35
T=50
T=100
F3Loc filter ( mono )
.212/2.32
.264/1.89
.435/0.41
.577/0.22
.730/0.17
PadLoc frozen ( mono )
.352/0.85
.423/0.47
.562/0.29
.620/0.26
.685/0.25
F3Loc filter ( comp_s )
.289/1.95
.337/1.34
.524/0.25
.687/0.13
.804/0.13
PadLoc frozen ( comp_s )
.420/0.54
.513/0.32
.687/0.19
.736/0.19
.737/0.19
Bonn per-sequence (columns T=150/300/450 )
Appendix
Table S13 : Unconditional-metric grid : SWA ↑ / all-trial median last-10 RMSE (m) ↓ for the runs behind the main paper’s SR cells. Top: Gibson, F3Loc depth sources (filter: per-frame readout; PadLoc: frozen). Bottom: Bonn per sequence, PadLoc cons with wheel and GT motion (frozen) vs UnLoc (own b^ , per-frame forward readout).
depth
T
n
SR@1 m
med/p90 ( ∘ )
wrong fam.
fail. wrong-fam.
UnLoc μ
15
268
0.731
2.6 / 90.1
13.8 %
37/72
UnLoc μ
20
200
0.840
1.7 / 13.1
9.5 %
19/32
UnLoc μ
35
111
0.919
1.7 / 7.3
6.3 %
7/9
UnLoc μ
50
78
1.000
1.3 / 3.5
0.0 %
–
UnLoc μ
100
37
1.000
1.4 / 3.6
0.0 %
–
comp_s
15
268
0.563
3.4 / 172.7
25.7 %
68/117
Appendix
Table S14 : Heading error and wall-family failures (Gibson cells, frozen readout; chunk-anchor heading vs. ground truth, wrapped to [0∘,180∘] ). Wrong family: error ≥45∘ . Last column: failed chunks that are wrong-family / all failed chunks.
scene
T=15
T=20
T=35
T=50
T=100
Spencerville
0.72 (39)
0.79 (29)
0.81 (16)
1.00 (11)
1.00 (5)
Spotswood
0.75 (20)
0.73 (15)
1.00 (8)
1.00 (6)
1.00 (3)
Springhill
0.74 (27)
0.90 (20)
0.91 (11)
1.00 (8)
1.00 (4)
Stilwell
0.71 (45)
0.85 (34)
0.89 (19)
1.00 (13)
1.00 (6)
Stokes
0.62 (24)
0.78 (18)
0.70 (10)
1.00 (7)
1.00 (3)
Sumas
0.95 (20)
0.93 (15)
1.00 (8)
1.00 (6)
1.00 (3)
Appendix
Table S15 : Per-scene SR@1 m and pooled Wilson 95 % intervals (UnLoc μ depth, PadLoc frozen; chunk counts in parentheses).