OmniVLoc: Topo-metric Visual Localization for Omnidirectional Cameras

Zhuo Yu; Shuyang Zhang; Huihua Zhu; Wenjia Ouyang; Zuo Zuo; Jianhao Jiao; Yilong Zhu; Xiaopin Zhong

cs.CVOct 20, 2026

OmniVLoc: Topo-metric Visual Localization for Omnidirectional Cameras

Authors: Zhuo Yu, Shuyang Zhang, Huihua Zhu, Wenjia Ouyang, Zuo Zuo, Jianhao Jiao, Yilong Zhu, Xiaopin Zhong,

Organizations: College of Mechatronics and Control Engineering, Shenzhen University, Shenzhen, Guangdong, 518060, China. · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, Guangdong, 518060, China. · College of Artificial Intelligence, Xi’an Jiaotong University, Xi’an, Shaanxi, 710049, China. · Robot Perception and Learning Lab, Intelligent Robotics, Department of Computer Science, University College London, Gower Street, WC1E 6BT, London, UK. ·

Abstract

Omnidirectional cameras provide surround-view information, significantly overcoming the field-of-view constraints for traditional visual localization systems. However, existing localization systems fail to fully exploit the characteristics of omnidirectional cameras, struggling to deliver accurate and robust localization performance amidst scene content variations and illumination-induced appearance changes. In this work, we propose OmniVLoc, a surround-view localization system based on a topo-metric framework. By integrating state-of-the-art global place recognition and monocular depth estimation methods, we introduce a multi-constrained visual localization paradigm. By solving a scale-aware pose graph and performing precise spherical bundle adjustment, we accurately recover the metric-level pose for each query frame.Furthermore, a hierarchical verification mechanism is designed to filter out erroneous outputs across different stages, effectively bolstering the overall system robustness. Extensive experiments on both public and our custom datasets validate the superior accuracy and reliability of OmniVLoc. To benefit the community, our source code and dataset will be made publicly available upon publication.

Figures & tables

Fig. 1: Comparison between our topo-metric map and traditional 3D feature map. Traditional 3D mapping relies on multi-view triangulation and suffers from geometric degeneracy and complex updates. Our topological map treats pose-annotated reference panoramas (orange) as independent database keys for the query (red), eliminating rigid geometric constraints, enabling effortless map management and continuous updates.

Fig. 2: Overview of our system. System front-end processes omnidirectional images to obtain monocular depth with scale uncertainty, global place candidates, and their feature associations with the query frame. The back-end utilizes a factor graph to eliminate scale uncertainty, followed by pose optimization with temporal constraints and bundle adjustment. Across the entire pipeline, we enforce strict consistency checks on intermediate outputs to avoid system degradation.

Fig. 3: Demonstration of our pose optimization process. The dashed lines illustrate the possible query positions yielded by a single PnP operation under unknown scales, while the solid lines indicate the valid candidates restricted by scale boundaries of $[0.5,2.0]$ .

Fig. 4: Demonstration of our temporal factor graph and our consistency verification.

Fig. 5: Our custom dataset and capturing devices. Plaza exhibits extreme day-to-night appearance shifts in an outdoor setting. Corridor involves a narrow indoor space with varying illumination. Avenue introduces repetitive foliage alongside ongoing construction events.

Method	ERP	IMU	Loc. Mode	Map Representation
Stella [ 30 ]	✔	✘	✔	2D&3D Features + BoW
HLoc [ 12 ]	✘	✘	✔	2D&3D Features + NetVLAD
ACE0 [ 17 ]	✘	✘	✔	Network Param.
Ours	✔	✔	✔	2D Features + MegaLoc

TABLE I: Supported Features of the Baseline Methods

Dataset	Scene	Seq.	Single-frame Localization					Sequence-based Localization
Dataset	Scene	Seq.	HLoc-Opt [ 12 ]	HLoc-Fus [ 12 ]	ACE0-Opt [ 17 ]	ACE0-Fus [ 17 ]	Ours-NoSeq	Stella [ 30 ]	Ours-Full
360Loc	Atrium	D1	2.966 / 0.346	2.510 / 0.707	1.304 / 0.780	1.583 / 1.168	2.613 / 0.321	0.320 / 0.282	0.079 / 0.054
		D2	2.096 / 0.912	1.698 / 0.879	1.866 / 0.999	2.556 / 1.472	0.344 / 0.096	1.049 / 0.908	0.231 / 0.111
		N1	2.743 / 1.017	3.120 / 1.147	- / -	- / -	2.071 / 0.193	- / -	0.407 / 0.166
		N2	0.988 / 0.984	1.007 / 0.914	- / -	- / -	0.322 / 0.143	0.956 / 0.957	0.309 / 0.155
	Concourse	D1	0.570 / 0.234	0.338 / 0.236	1.406 / 0.746	1.603 / 0.940	0.030 / 0.019	2.116 / 1.442	0.029 / 0.018

TABLE II: Localization accuracy on the 360Loc and custom datasets. Each entry reports mean/median ATE. ’-’ represents initialization failure or ATE more than

10

Panel of Figure 8 — Fig. 6: Qualitative cases of our multiple reference localization. The text overlays indicate the retrieved map node ID, descriptor similarity, number of RANSAC inliers, and the single-frame PnP pose error, with green/red denoting accepted/rejected candidates by our verification module.

Fig. 7: Visualization of map storage at 360Loc Atrium scene. Note that the map components are sorted by size and the x-axis is in logarithmic scale.

$N_{\mathrm{ret}}$	Time Consumption (ms)				ATE		Avg. Map-ref Usage per Query
$N_{\mathrm{ret}}$	Frontend	Sph. PnP	BA	Total	RMSE	Median	Avg. Map-ref Usage per Query
1	300.3	109.8	20.0	483.7	1.434	1.258	0.98
3	410.3	121.3	31.3	618.1	0.096	0.057	2.21
5	521.7	140.4	49.8	765.9	0.079	0.054	3.23
8	685.2	166.9	73.6	979.6	0.076	0.055	4.42
10	798.2	201.6	76.7	1130.9	0.076	0.058	4.98

TABLE III: Comparison of top-

N_{\mathrm{ret}}

Retrieval on computation cost, localization accuracy, and map-reference usage at 360Loc Atrium D1.

Robot Localization

Robot localization estimates a robot’s position and orientation from onboard sensors, maps, and observations of its surroundings.

See all 179 papers

Topometric Autonomous Vehicle Localization by Combining Visual Embeddings and Feed-Forward 3D Models
Aug 6, 2026Eulogio Quemada-Torres, Alberto Jaenal, Francisco-Angel Moreno +1Visual Place Recognition Map-Based Localization
Machine Perception and Intelligent Robotics Group (MAPIR), Department of Systems Engineering and Automation, Malaga Institute for Mechatronics Engineering and Cyber-Physical Systems (IMECH.UMA), University of Malaga, Malaga, Spain · Robotics, Computer Vision and Artificial Intelligence Group (RoPeRT), Aragon Institute for Engineering Research (I3A), University of Zaragoza, Zaragoza, Spain
MultiLoc: Look Around As You Localize for Fast And Robust Visual Re-localization
Mar 28, 2026Nobel Dang, Bing LiRelative Pose Estimation Camera Pose Estimation
Clemson University, SC, USA
RHO: Robust Holistic OSM-Based Metric Cross-View Geo-Localization
Mar 29, 2026Junwei Zheng, Ruize Dai, Ruiping Liu +7Cross-View Geo-Localization Map-Based Localization
Karlsruhe Institute of Technology · ETH Zurich · UCL +2

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

Topometric Autonomous Vehicle Localization by Combining Visual Embeddings and Feed-Forward 3D Models

Eulogio Quemada-Torres, Alberto Jaenal, Francisco-Angel Moreno, Javier Gonzalez-Jimenez

Machine Perception and Intelligent Robotics Group (MAPIR), Department of Systems Engineering and Automation, Malaga Institute for Mechatronics Engineering and Cyber-Physical Systems (IMECH.UMA), University of Malaga, Malaga, Spain · Robotics, Computer Vision and Artificial Intelligence Group (RoPeRT), Aragon Institute for Engineering Research (I3A), University of Zaragoza, Zaragoza, Spain

Visual Place Recognition Map-Based Localization Robot Localization

Open paper AskPDF arXiv

Effective Visual Localization (VL) requires a map of the environment that combines compactness for efficient scalability with robustness against visual appearance changes and metric precision. Through low-dimensional image embeddings, Visual Place Recognition (VPR) is able to successfully meet the first two requirements, but its low metric accuracy makes it less suitable than standard VL approaches based on local features or neural representations. This limitation can be overcome by integrating VPR with the accurate local trajectory estimates produced by feed-forward neural 3D geometry (FF3D) models. In this paper, we address sequential appearance-based localization through a topometric framework that iteratively combines probabilistic VPR with FF3D metric pose estimation in controlled image sets. Our approach proposes an automatic offline mapping tool that models the topometric pose-appearance interaction in the different parts of the scene. This map is later employed by an online particle filter that estimates the pose from odometry and belief over places for FF3D inference, successfully incorporating neural metric estimation into probabilistic appearance-based localization. We extensively evaluate the framework on three known benchmarks, demonstrating substantial improvements over existing appearance-based methods. The modularity of our approach allows the descriptor extractor and FF3D model to remain interchangeable, and a focused analysis further shows that sequential belief can mitigate severe failures under perceptual aliasing.

Abstract

Robot Localization

Explore similar work

Topometric Autonomous Vehicle Localization by Combining Visual Embeddings and Feed-Forward 3D Models

MultiLoc: Look Around As You Localize for Fast And Robust Visual Re-localization

RHO: Robust Holistic OSM-Based Metric Cross-View Geo-Localization