Recovering from Pose-Selection Failures in Repeated Visual Localization

Jaehyun Kim; Wonseok Kang; Tae-Wan Kim

cs.CVOct 12, 2026

Recovering from Pose-Selection Failures in Repeated Visual Localization

Authors: Jaehyun Kim, Wonseok Kang, Tae-Wan Kim

Organizations: Department of Naval Architecture and Ocean Engineering, Seoul National University, Seoul, South Korea

Abstract

Repetitive structures can give an incorrect pose candidate more inlier support than an accurate one, causing visual localization to select it. History-Guided Pose Reselection (HPR) uses coarse position forecasts from up to five past localization outputs to distinguish well-separated current-query candidates. The best-fitting forecast proposes an alternative; HPR accepts it only if every well-fitting forecast is much closer to it than to the localizer's pose. HPR needs no odometry and returns the selected candidate unchanged, using prediction for selection rather than pose refinement. With one development-set threshold, HPR improves recall within 5 meters and 10 degrees in all 12 main configurations of Hierarchical Localization (HLoc), MASt3R-based pose generation, and JIST-based retrieval, losing successes on fewer than 0.4 percent of queries. On NAVER Gangnam B1, HPR corrects 97 HLoc failures and loses six successes; gains persist within 0.25 meters and 2 degrees on all three HLoc test sets. Median HPR runtime on HLoc candidates is below one millisecond, excluding pose generation. Recovery weakens under less frequent localization.

Figures & tables

Fig. 1: HPR overview on an HLoc B1 query. (1–2) The best-fitting forecast ( $\star$ ) proposes the nearest alternative. (3) Retained forecasts meet the 5.3% distance-ratio limit ( $\tau\approx 0.899$ ). (4) HPR returns the alternative unchanged; history stores the baseline. Filled/open triangles mark baseline history/forecasts; labels 0–4 identify omitted observations. Region A is enlarged; distances are 3-D. Thumbnails show each group’s top-retrieved reference. Images, LiDAR, and ground truth (GT) are visual context, not HPR inputs.

Dataset	Queries	Streams	Median span
Gangnam B2 [ 4 ]	916	6	8 s
Gangnam B1 [ 4 ]	2,620	8	8 s
360Loc night [ 28 ]	6,982	12	4 frames
Hyundai [ 4 ]	1,751	4	8 s
OpenLORIS Corridor [ 29 ]	786	4	2 s
LaMAR [ 30 ]	2,732	19	4 s

TABLE I: Query Streams and Five-Position History Spans

	Baseline successes		Recoverable	HPR $C/H$
Configuration	Primary	Strict	$A/F$	Primary	Strict
HLoc B1	2,200	1,886	210/420	97/6	66/5
HLoc night	6,037	3,747	121/945	76/1	18/0
HLoc Hyundai	1,714	1,469	21/37	16/0	14/0
MASt3R B1	2,400	269	69/220	38/3	4/0
MASt3R night	3,813	65	440/3,169	16/3	0/0

TABLE II: Pose Recovery Across Configurations

		B2 (dev.)	B1		360Loc night		Hyundai
Method	$\tau$	Primary	Primary	Strict	Primary	Strict	Primary	Strict
Baseline	–	$66.92\ (0/0)$	$83.97\ (0/0)$	$0/0$	$86.47\ (0/0)$	$0/0$	$97.89\ (0/0)$	$0/0$
Confidence (max.)	–	$77.40\ (109/13)$	$\mathbf{88.55}\ (131/11)$	$99/6$	$87.01\ (67/29)$	$37/21$	$97.49\ (15/22)$	$12/22$
Confidence (cal.)	0.1512	$76.31\ (91/5)$	$88.24\ (115/3)$	$90/1$	$86.94\ (48/15)$	$22/7$	$97.60\ (14/19)$	$11/20$
Inlier ratio	0.8376	$66.59\ (2/5)$	$82.52\ (2/40)$	$1/27$	$85.49\ (0/68)$	$1/24$	$97.09\ (2/16)$	$2/6$
Constant position	0.9846	$67.90\ (14/5)$	$84.50\ (18/4)$	$9/4$	$86.88\ (29/0)$	$8/0$	$98.00\ (3/1)$	$2/0$

TABLE III: HLoc Pose Selection at Primary and Strict Criteria

Fig. 2: Correction–error trade-offs at 5 m/

10^{\circ}

. Markers show operating points selected under increasing B2 error limits and transferred unchanged to targets. HMMs use

\sigma=8

m; stars and filled crosses mark HPR and row-normalized HMM at limit 5. The night inset enlarges low-error points.

Setting	B1	Night	Hyundai
Default HPR	$97/6$	$76/1$	$16/0$
Window: 4 positions	$68/4$	$62/0$	$6/0$
Window: 7 positions	$103/18$	$70/7$	$17/1$
Residual factor: 1	$66/3$	$69/2$	$8/0$
Residual factor: 3	$100/6$	$78/1$	$16/0$
Time/frame indexing	$99/7$	$76/1$	$13/0$

TABLE IV: HLoc Sensitivity, Candidate Screening, and Calibration Uncertainty

Fig. 3: All 6,697 HLoc test-set proposals that would change primary success status.

d_{a}

and

d_{b}

are the alternative’s and baseline’s distances to the minimum-margin forecast. The shaded region below

d_{a}\approx 0.053d_{b}

satisfies

G_{t}\geq\tau

; rings mark the seven introduced errors. The inset includes out-of-range proposals, with six corrections from extreme baselines (screening: Table IV (b)).

	Confidence (cal.)		Confidence + HPR
			Unchanged $\tau$		Recalibrated $\tau$
Configuration	P	S	P	S	P	S
HLoc B1	115/3	90/1	164/3	116/1	138/3	103/1
HLoc night	48/15	22/7	89/4	32/6	69/7	28/6
HLoc Hyundai	14/19	11/20	16/7	13/8	14/17	11/18
MASt3R	21/1,532	0/190	21/1,522	0/189	21/1,532	0/190

TABLE V: Confidence Selection and History-Based Recovery

Scene, device	$A/F$	Primary	Strict	LaMAR criteria
		5 m/ $10^{\circ}$	0.25 m/ $2^{\circ}$	1 m/ $5^{\circ}$	0.1 m/ $1^{\circ}$
CAB phone	13/77	4/0	2/0	2/0	1/0
CAB HL	84/108	9/0	5/0	9/0	2/0
HGE phone	10/33	1/0	1/0	1/0	1/0
HGE HL	63/192	2/12	0/8	2/12	0/4
LIN phone	2/18	0/0	0/0	0/0	0/0

TABLE VI: Fixed-Threshold Transfer to LaMAR: Camera-Image

C/H

	B1		Night		Hyundai
Stride	P	S	P	S	P	S
1	97/6	66/5	76/1	18/0	16/0	14/0
2	63/6	42/6	65/0	18/0	13/0	11/0
4	25/10	14/8	48/1	11/0	3/0	2/0
8	15/11	10/9	40/0	11/0	5/0	4/0
Past GT	167/0	105/0	90/0	22/0	16/0	14/0

TABLE VII: HLoc Recovery Under Sparse Updates and Past-GT History

Map-Based Localization

Map-Based Localization estimates a device or robot’s position and orientation by matching sensor observations to stored maps.

See all 106 papers

Reference-Induced Consensus for Selective Posed-Reference Visual Localization
Jul 6, 2026Wonseok Kang, Jaehyun Kim, Jeongmin Lee +1Camera Pose Estimation Robot Localization
DeepFine Co., Ltd., Seoul, South Korea · Seoul National University, Department of Naval Architecture and Ocean Engineering, Seoul, South Korea
Hierarchical Visual Relocalization with Nearest View Synthesis from Feature Gaussian Splatting
Mar 31, 2026Huaqi Tao, Bingxi Liu, Guangcheng Chen +3Camera Pose Estimation
Southern University of Science and Technology, Shenzhen, China. · Institute of Automation, Chinese Academy of Sciences, Beijing, China.
PoseCompass: Intelligent Synthetic Pose Selection for Visual Localization
May 12, 2026Yanan Zhou, Zhaoyan Qian, Yanli Li +3Visual Place Recognition Camera Pose Estimation
The University of Sydney, Australia · University of St Andrews, United Kingdom

Jul 6, 2026J/K move · Enter open · S save

Reference-Induced Consensus for Selective Posed-Reference Visual Localization

Wonseok Kang, Jaehyun Kim, Jeongmin Lee, Tae-Wan Kim

DeepFine Co., Ltd., Seoul, South Korea · Seoul National University, Department of Naval Architecture and Ocean Engineering, Seoul, South Korea

Camera Pose Estimation Robot Localization

Open paper AskPDF arXiv

We present RIC-Loc (Reference-Induced Consensus localization), a scene-training-free posed-reference localizer that is SfM-point-map-free in its main estimator: it uses known reference poses, but not precomputed SfM 3D map points, query-to-map 2D-3D matches, or query-to-map PnP. A frozen VGGT pass predicts local camera poses, depth, and query-reference tracks for a query and selected references. Each reference induces one map-frame SE(3) query-pose hypothesis, robust consensus estimates the pose, and the preserved hypothesis structure yields two reliability scores: spatial dispersion and a track-conditioned covariance score. On the covariance-eligible set, the two scores are complementary for held-out, ground-truth-free failure detection across indoor, outdoor, and large-scale low-texture benchmarks: the joint policy is strongest in textured scenes and the covariance score in the low-texture regime, and the hypothesis-derived scores consistently outperform the standard retrieval-score gap and random rankings. Without per-scene training the consensus estimator remains accurate -- competitive with structure-based localization indoors and improving over a comparable feed-forward baseline -- giving an effective selective operating regime for posed-reference localization. Code is available at https://github.com/SNU-DLLAB/ric_loc.

Abstract

Map-Based Localization

Explore similar work

Reference-Induced Consensus for Selective Posed-Reference Visual Localization

Hierarchical Visual Relocalization with Nearest View Synthesis from Feature Gaussian Splatting

PoseCompass: Intelligent Synthetic Pose Selection for Visual Localization