TNVD-SC: Diagnosing Scan Context Failures in Forests with Terrain-Normalized Vertical Distributions

Jineon Kim; Jhonghyun An

cs.CVOct 19, 2026

TNVD-SC: Diagnosing Scan Context Failures in Forests with Terrain-Normalized Vertical Distributions

Authors: Jineon Kim, Jhonghyun An

Organizations: School of Computing, Gachon University, Seongnam 13120, Republic of Korea.

Abstract

Long-term LiDAR place recognition in forests is difficult because repetitive vegetation and uneven terrain create severe structural aliasing. We diagnose Scan Context (SC) by separating candidate retrieval from fine matching: on the standardized TreeLoc++ Wild-Places protocol, a TNVD support key places a valid location in the top-25 set for approximately 99% of queries, while conventional SC matching on those candidates reaches only 61.84% mean R@1. With an identical raw-SC candidate set, terrain normalization improves max-height cells by only 1.21 points but vertical-distribution cells by 19.72 points. These observations motivate Terrain-Normalized Vertical Distribution Scan Context (TNVD-SC), which preserves the SC ring-sector topology while encoding terrain-relative vertical probability distributions. Under the standardized TreeLoc++ protocol, TNVD-SC reaches 91.83% mean R@1 and attains the highest R@1 mean and R@1 stability among the compared methods, while TreeLoc++ remains stronger in mean MR, MF1, and AUC. The unchanged method reaches 85.14% mean R@1 on the four-session CVTNet-style NCLT evaluation. A lossless count-coded implementation with exact hierarchical Jensen-Shannon search reduces descriptor payload by 3.86x and compute latency by 23.8% without changing the audited retrieval outputs.

Figures & tables

Fig. 1: Overview of TNVD-SC. Query and database scans share the TNVD backbone. A 480-D support-only key retrieves top- $K$ candidates, while the ordered full TNVD tensor performs circular fine matching. Bottom: conventional SC stores one maximum height per cell, whereas TNVD retains its terrain-relative vertical distribution.

Fig. 2: Candidate-budget diagnosis for conventional SC. Candidate recall continues to rise with $K$ , whereas final SC R@1 saturates. The single dashed line marks the $K=25$ operating point used by TNVD-SC.

Retrieval	Matcher	Venman		Karawatha		Mean
		R@1	R@25	R@1	R@25	R@1
SC ring key	SC	40.53	76.32	53.37	86.82	46.95
SC ring key	TNVD	68.72	76.32	79.15	86.82	73.94
TNVD support	SC	57.20	98.89	66.47	98.76	61.84
TNVD support	TNVD	92.24	98.89	91.41	98.76	91.83

TABLE I: Component decomposition on standardized Wild-Places. R@25 is candidate recall before fine reranking.

Cell representation	Terrain norm.	Venman	Karawatha	Mean
Max-height	No	40.53	53.37	46.95
Max-height	Yes	42.65	53.67	48.16
Vertical distribution	No	50.66	57.78	54.22
Vertical distribution	Yes	68.72	79.15	73.94

TABLE II: Terrain normalization under an identical raw-SC Top-25 candidate set (pair-mean R@1, %).

Fig. 3: Effect of terrain normalization on the vertical representation. The same local forest corridor is shown in absolute-height and terrain-relative coordinates. Removing low-frequency terrain relief aligns the local vertical reference while preserving vegetation-layer structure, which is subsequently encoded by TNVD.

Category	Method	Venman01–04 (12 query–database pairs)								Karawatha01–04 (12 query–database pairs)
		R@1		MR		MF1		AUC		R@1		MR		MF1		AUC
		Mean	$S$	Mean	$S$	Mean	$S$	Mean	$S$	Mean	$S$	Mean	$S$	Mean	$S$	Mean	$S$
Learning	TransLoc3D	.553	.464	.008	-.048	.707	.678	.604	.511	.711	1.316	.013	-.103	.831	1.561	.756	1.268
	LoGG3D-Net	.660	1.242	.003	-.282	.813	1.382	.732	.872	.897	1.848	.002	-.278	.948	2.299	.921	1.638
	MinkLoc3Dv2	.755	.782	.033	.161	.857	1.030	.818	.881	.849	1.607	.062	.150	.919	1.889	.911	1.592
	BEVPlace++	.735	.701	.091	-.025	.840	.920	.841	.842	.767	.934	.016	-.193	.868	1.179	.767	.822

TABLE III: Inter-session place recognition performance on Wild-Places under the standardized TreeLoc++ Table-5 protocol. Each site contains 12 directed query–database pairs. For each metric, we report the pair mean and stability ratio

S

from Eq. ( 12 ). Published baselines are from TreeLoc++ [ 7 ] . Best values are bold.

Method	02-05	06-15	02-23	04-05	Mean
PointNetVLAD	74.6	61.2	46.9	44.9	56.90
MinkLoc3D	80.2	63.0	50.7	48.2	60.53
OverlapTransformer	86.1	63.9	53.6	49.6	63.30
SeqOT	91.7	76.2	69.1	63.9	75.23
CVTNet	93.2	79.3	81.7	78.0	83.05
SC (ours)	91.89	69.09	70.15	63.77	73.73

TABLE IV: NCLT four-session CVTNet evaluation (R@1, %). Published baselines are from CVTNet [ 21 ] ; our rows use the released full-database evaluator.

Sequence	SC	TNVD-SC	TreeLoc++	$\Delta$
Evo	30.49	40.78	93.90	+10.29
Forest of Dean	22.39	73.17	76.27	+50.78
Stein am Rhein	35.44	45.42	93.69	+9.98
Wytham	8.16	68.37	39.80	+60.21
Mean	24.12	56.94	75.92	+32.82

TABLE V: Oxford Forest intra-session R@1 (%) under the common 5 m / 50-frame protocol. TreeLoc++ is re-run on the same released sequences and evaluated-query cohorts.

Ablation	Setting	Venman	Karawatha	Mean
Retrieval key	Mean, 480-D	89.34	89.59	89.47
	Support only, 480-D	92.24	91.41	91.83
	Mean+Std, 960-D	88.04	87.82	87.93
	Mean+Support, 960-D	92.31	91.40	91.86
	Mean+Std+Support, 1440-D	92.37	91.36	91.87
Similarity	Cosine	92.38	91.30	91.84

TABLE VI: Retrieval-key, similarity, and representation ablations on standardized Wild-Places (R@1, %). The full-distribution cosine entry intentionally repeats the cosine-similarity row as the reference condition for the binary-representation ablation.

Setting	Value	Key dim.	Venman	Karawatha / Mean
$B$ ( $H=18$ m)	6	120	81.66	81.87 / 81.77
	12	240	89.21	88.60 / 88.91
	24	480	92.24	91.41 / 91.83
	36	720	93.34	92.39 / 92.87
$H$ ( $B=24$ )	12 m	480	92.32	89.90 / 91.11
	18 m	480	92.24	91.41 / 91.83

TABLE VII: Vertical discretization sensitivity on standardized Wild-Places (R@1, %).

Implementation	Payload [kB]	DB [MiB]	Query [ms]	Rate [Hz]
SC reference	9.7	–	18.73	53.4
TNVD float exhaustive	239.0	504.9	228.55	4.38
TNVD count + exact hier.	61.9	188.5	174.27	5.74

TABLE VIII: Implementation efficiency. SC timing is reported as a reference from the same

K=25

, 500-query clean benchmark; its DB-RSS value was not recorded with the final compact-memory campaign.

Fig. 4: Long-range TNVD-SC alias. A 159.0 m false candidate outranks the 1.69 m GT; regions A/B explain 95.8% of the net similarity gap. The lower row collapses the 24-bin TNVD tensor into three height bands (0–6, 6–12, and 12–18 m) for display only; matching uses all $B=24$ bins.

Fig. 5: SC failures recovered by TNVD-SC. P50/P90 examples use representative pairs nearest to each site’s mean gain; insets show the 5 m GT region.

LiDAR Place Recognition

LiDAR place recognition retrieves previously visited locations from 3D scans to support robot localization and SLAM loop closure.

See all 20 papers

UpDown-SC: Gravity-Canonicalized Dual-Envelope Scan Context for Indoor LiDAR Place Recognition
Sep 24, 2026Jie Xu, Yongxin Yang, Ziyi Jin +4LiDAR Place Recognition LiDAR Localization
Anyverse Dynamics, Beijing, China
TreeLoc++: Robust 6-DoF LiDAR Localization in Forests with a Compact Digital Forest Inventory
Mar 4, 2026Minwoo Jung, Dongjae Lee, Nived Chebrolu +3Robotics LiDAR Localization
Department of Mechanical Engineering, Seoul National University, Seoul, South Korea · Department of Computer Science and Engineering, Indian Institute of Technology Bombay, India · Oxford Robotics Institute, Department of Engineering Science, University of Oxford, Oxford, UK
Heterogeneous LiDAR Early Fusion and Learned Re-Ranking Strategy for Robust Long-Term Place Recognition in Unstructured Environments
Jun 11, 2026Judith Vilella-Cantos, Juan José Cabrera, Mónica Ballesta +2LiDAR Place Recognition Agricultural Robotics
University Institute for Engineering Research, Miguel Hernández University, Avda. de la Universidad s/n, Edificio Innova, Elche, 03202, Alicante, Spain

Sep 24, 2026J/K move · Enter open · S save

UpDown-SC: Gravity-Canonicalized Dual-Envelope Scan Context for Indoor LiDAR Place Recognition

Jie Xu, Yongxin Yang, Ziyi Jin, Kangjin Yu+3

Anyverse Dynamics, Beijing, China

LiDAR Place Recognition LiDAR Localization Map-Based Localization Robot Localization

Open paper AskPDF arXiv

LiDAR place recognition is a key front end for loop closure and global relocalization, yet indoor retrieval remains difficult when attitude or sensor mounting height changes between mapping and query sessions. Scan Context stores the maximum height in each polar cell; indoors, broad ceilings can suppress the lower and mid-level geometry that distinguishes adjacent rooms and corridors. We present UpDown-SC, a training-free polar descriptor that first canonicalizes gravity and then represents two complementary surfaces: the upper envelope of lower/middle structures and the lower envelope of overhead structures. Their physical split is estimated once from a cell-balanced map height distribution and reused by every query. A mask-aware, non-uniform two-channel distance retains discriminative lower-level evidence while limiting sensitivity to its cross-session variation, without treating unobserved cells as zero-height measurements. Conventional Scan Context shortlisting and circular yaw alignment are retained, so retrieved hypotheses directly initialize geometric verification. Experiments across repeated indoor sessions, mounting-height changes, mixed outdoor-to-indoor trajectories, and an outdoor transfer sequence show more reliable first-choice retrieval on the indoor and mounting-height-varied sessions. A paired test finds a significant gain over Scan Context on the in-house sessions. UpDown-SC also gives the best or second-best F1max and AUPR under threshold-based acceptance while retaining a lightweight CPU front end. Continuous replay confirms that the retrieved hypotheses support metric prior-map localization. Code and evaluation artifacts: https://github.com/jiejie567/updown-sc.

Abstract

LiDAR Place Recognition

Explore similar work

UpDown-SC: Gravity-Canonicalized Dual-Envelope Scan Context for Indoor LiDAR Place Recognition

TreeLoc++: Robust 6-DoF LiDAR Localization in Forests with a Compact Digital Forest Inventory

Heterogeneous LiDAR Early Fusion and Learned Re-Ranking Strategy for Robust Long-Term Place Recognition in Unstructured Environments