Scene Representation using 360° Saliency Graph and its Application in Vision-based Indoor Navigation

Preeti Meena; Himanshu Kumar; Sandeep Yadav

cs.CVMar 20, 2026

Scene Representation using 360° Saliency Graph and its Application in Vision-based Indoor Navigation

Authors: Preeti Meena, Himanshu Kumar, Sandeep Yadav

Organizations: Discipline of Electrical Engineering, Indian Institute of Technology Jodhpur, Jodhpur-342030, India

Abstract

A Scene, represented visually using different formats such as RGB-D, LiDAR scan, keypoints, rectangular, spherical, multi-views, etc., contains information implicitly embedded relevant to applications such as scene indexing, vision-based navigation. Thus, these representations may not be efficient for such applications. This paper proposes a novel 360° saliency graph representation of the scenes. This rich representation explicitly encodes the relevant visual, contextual, semantic, and geometric information of the scene as nodes, edges, edge weights, and angular position in the 360° graph. Also, this representation is robust against scene view change and addresses challenges of indoor environments such as varied illumination, occlusions, and shadows as in the case of existing traditional methods. We have utilized this rich and efficient representation for vision-based navigation and compared it with existing navigation methods using 360° scenes. However, these existing methods suffer from limitations of poor scene representation, lacking scene-specific information. This work utilizes the proposed representation first to localize the query scene in the given topological map, and then facilitate 2D navigation by estimating the next required movement directions towards the target destination in the topological map by using the embedded geometric information in the 360° saliency graph. Experimental results demonstrate the efficacy of the proposed 360° saliency graph representation in enhancing both scene localization and vision-based indoor navigation.

Figures & tables

Panel of Figure 1 — Fig. 1: (A) Orientation-rich $360^{\circ}$ saliency graph representation of a scene, constructed from an RGB-D panorama image. (B) Localization in a topological map and positioning within a localized scene using $360^{\circ}$ saliency graphs. Localization: via scene graph matching using visual, semantic, and contextual information embedded in the $360^{0}$ saliency graph; Positioning in the scene: by aligning geometric information in a $360^{0}$ saliency graph.

Panel of Figure 2 — Fig. 2: Overall framework of the proposed method for Queryable Topological Map Generation , and Scene Localization . For Queryable Topological Map Generation , a building is processed to extract the floor map (a 2D layout of the building), the region semantic mask (a segmentation mask labeling larger semantic regions such as rooms and hallways), and the object semantic mask (a finer-grained segmentation mask labeling individual objects) to construct a queryable topological map $T$ that facilitates the navigation path/topological connection between the scene locations $\mathcal{V}$ (nodes in red and scene transition nodes in blue). At each scene location in $\mathcal{V}$ , an orientation-rich $360^{\circ}$ saliency graph $G$ is generated from an RGB-D panorama image as a scene representation. In Scene Localization , a query panoramic RGB-D image of a scene is localized within the topological map $T$ . For this, the query $360^{\circ}$ saliency graph is constructed from the query image and matched against the precomputed saliency graphs $G$ associated with the map nodes to extract the location with the highest similarity above a minimum threshold. Afterwards, positioning is performed by aligning the $360^{\circ}$ saliency graphs of query and reference scenes to estimate the directions for navigation towards the target destination on the topological map.

Fig. 3: Alignment of query graph $G^{q}$ and matched candidate graph $\mathcal{G}_{m}$ by estimating shift $\Delta$ and orientation using underlying geometry of reference points ( $c$ , $c^{\prime}$ ) and matched salient objects.

Panel of Figure 4 — Fig. 4: Sample panorama images for different viewer positions and orientation difference ( $\theta_{e}$ ).

# of instances	$E_{p}$	$E_{\theta}$
$105$	$0.6995$ m	$16.59^{0}$

TABLE I: Mean position (

E_{p}

) and orientation (

E_{\theta}

) errors, computed between shifted and original/ground-truth poses.

Fig. 5: An example of navigation on topological map. First, determine current orientation $\theta_{c}$ at the current location $(X_{c},Y_{c})$ using the saliency graph after performing localization. Next, the rotation angle $\theta_{r}$ between the direction of the topological path and $\theta_{c}$ is calculated after performing alignment using positioning. The navigation algorithm then adjusts the heading based on this $\theta_{r}$ , ensuring accurate movement towards the next target location $(X_{t},Y_{t})$ on the shortest path algorithm.

Methods	MMA@3	Methods	MMA@3
SP [ 50 ] +NN	$32.43$	QuadT [ 55 ]	41.72
SP [ 50 ] +SG [ 53 ]	$29.95$	LoFTR [ 56 ]	36.07
ASLFeat [ 52 ] +NN	$35.79$	COTR [ 57 ]	$46.07$
ALike [ 51 ] +NN	$34.55$	GT+SGAM_COTR [ 58 ]	$49.82$
ASpan [ 54 ]	$37.25$	Proposed	$\mathbf{87.8}$

TABLE II: Localization performance comparison in terms of MMA

@3

with state-of-the-art feature matching-based methods.

Methods	LDL [ 59 ]	FGPL [ 60 ]	MiC [ 43 ]	MFP-CNN [ 61 ]	SOVS [ 17 ]	Proposed
Acc ( $\uparrow$ )	$0.35$	$0.71$	$0.78$	$0.80$	$0.82$	$\mathbf{0.86}$

TABLE III: Comparison of mean accuracy for scene localization.

Panel of Table 9 — TABLE IV: Sample results for scene localization.

Method	$f$	$2f$	$3f$
SOVS [ 17 ]	$0.47$	$0.36$	$0.29$
Proposed	$0.65$	$0.50$	$0.42$

TABLE V: Robustness of scene localization accuracy under varied focal length.

Fig. 6: An example of proposed localization + MiC [ 43 ] navigation method.

Methods	Navigation Acc.
	OSR( $\uparrow$ )	SR( $\uparrow$ )	SPL( $\uparrow$ )
FAST-MATTN [ 39 ]	$28.20$	$14.40$	$7.19$
CKR [ 15 ]	$31.44$	$19.14$	$11.84$
Airbert [ 40 ]	$34.51$	$27.89$	$21.88$
HOP [ 41 ]	$36.24$	$31.78$	$26.11$
DUET [ 42 ]	$51.07$	$46.98$	$33.73$

TABLE VI: Impact of proposed localization method on navigation performance.

Panel of Figure 13 — Fig. 7: An example showing the room-type prediction using MiC [ 43 ] and proposed localization method.

Methods	Navigation Acc.
	SR( $\uparrow$ )	SPL( $\uparrow$ )
MiC [ 43 ]	${56.97}$	${43.60}$
Proposed Navigation	$\mathbf{76.92}$	$\mathbf{72.85}$

TABLE VII: Performance comparison of proposed navigation framework against MiC [ 43 ] .

Methods	Image	Objects	Salient objects	USG	$360^{\circ}$ saliency graph
Acc ( $\uparrow$ )	$0.68$	$0.76$	$0.82$	$0.84$	$\mathbf{0.86}$

TABLE VIII: Impact of different components on localization. (USG: unweighted saliency graph)

Method	without $360^{0}$ saliency graph	with $360^{0}$ saliency graph
SR	$58.73$	$\mathbf{76.92}$

TABLE IX: Impact of

360^{0}

saliency graph-based positioning on navigation performance.

Table 17

Method	Localiza-	Navigation Acc.
	tion Acc( $\uparrow$ )	SR( $\uparrow$ )	SPL( $\uparrow$ )
without perturbation	$0.86$	$76.92$	$72.85$
with spatial perturbation	$0.82$	$71.43$	$63.68$
with orientation perturbation	$0.74$	$68.75$	$62.05$
with object perturbation ( $50$ %) $P_{1}$	$0.78$	$69.71$	$61.21$
with object perturbation ( $66$ %) $P_{2}$	$0.67$	$40.28$	$31.77$

TABLE XII: Perturbation Analysis.

3D Scene Representation

3D scene representation studies how to encode geometry, appearance, semantics, and motion for reconstruction, rendering, generation, and interaction.

See all 242 papers

A Sparse Topometric Graph for Scalable Cross-Session Indoor Navigation
Oct 22, 2026Eun Sun Lee, Minkyun Seo, Gwangtak Bae +4Indoor Robot Navigation Graph Matching
Seoul National University
A Topological Representation with Object-Path Graphs for Open-Vocabulary Instance Navigation
Sep 21, 2026Linwei Zheng, Daojie Peng, Bingtao Wang +2Robot Navigation Open-Vocabulary Object Navigation
The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology
LEXI-SG: Monocular 3D Scene Graph Mapping with Room-Guided Feed-Forward Reconstruction
May 13, 2026Christina Kassab, Hyeonjae Gil, Matías Mattamala +2Simultaneous Localization and Mapping Monocular 3D Reconstruction
Department of Engineering Science, University of Oxford, UK · Department of Mechanical Engineering, Seoul National University, South Korea · School of Informatics, University of Edinburgh, UK

Oct 22, 2026J/K move · Enter open · S save

A Sparse Topometric Graph for Scalable Cross-Session Indoor Navigation

Eun Sun Lee, Minkyun Seo, Gwangtak Bae, Junho Kim+3

Seoul National University

Indoor Robot Navigation Graph Matching

Open paper AskPDF arXiv

A conventional embodied agent in an indoor environment builds a map by accumulating noisy real-world measurements and results from heavy inference to guide navigation. While the detailed measurements may differ due to scene changes or sensor settings, the unique topological structure of navigable passages or the connectivity between rooms may provide training-free guidance that complements local observations. We propose a sparse topometric graph as an efficient structural representation for robust indoor navigation, alleviating the need for fine-grained matching, which is computationally expensive, memory-intensive, and prone to real-world variations. The map consists of nodes that encode local 3D geometry and distance-weighted edges that capture traversable connectivity. A learned graph-matching network finds consistent subgraph structures across sessions by considering both node-level geometric descriptors and neighborhood connectivity. The holistic contextual matching is robust to local ambiguities caused by weak distinctiveness or repetitive structure, and establishes stable correspondences despite local scene changes, different sensor modalities, or limited spatial overlap. We analyze the practicality of the graph through cross-session point-goal navigation, where a goal specified in a prior reference map must be reached from an unknown initial pose in a later session. Experiments show that our method achieves the highest success rates (77.1% on Gibson and 40.0% under 3D-FRONT scene changes), while reducing memory by more than two orders of magnitude compared to point-level representations. The proposed approach further supports cross-session alignment and goal-directed navigation with strong scalability and robust real-world transfer.

Abstract

3D Scene Representation

Explore similar work

A Sparse Topometric Graph for Scalable Cross-Session Indoor Navigation

A Topological Representation with Object-Path Graphs for Open-Vocabulary Instance Navigation

LEXI-SG: Monocular 3D Scene Graph Mapping with Room-Guided Feed-Forward Reconstruction