Visual Floorplan Localization (FLoc)

• Visual Floorplan Localization is the task of determining a camera's 2D position and orientation within known indoor floorplans using visual observations and map cues.
• It employs probabilistic inference over discretized pose spaces and uses techniques like ray-based depth matching and cross-modal alignment to resolve pose ambiguities.
• Recent advancements integrate uncertainty modeling, semantic disambiguation, and 3D reconstruction-based proxies to enhance localization accuracy and robustness.

Visual Floorplan Localization (FLoc) is the problem of estimating where a visual observation was captured within a known indoor floorplan. In the formulations represented here, the target is typically a planar pose $S=(x,y,\theta)$ or $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$, with $(x,y)$ denoting 2D position in the map and $\theta$ or $\phi$ denoting yaw. The input may be a single RGB image, an image sequence with relative motion, or a stationary scan of posed images, together with a 2D floorplan that may encode occupancy geometry alone or richer semantics such as doors and windows. The appeal of FLoc follows from properties repeatedly emphasized in the literature: floorplans are lightweight, readily available, long-term persistent, and robust to changes in visual appearance; the central difficulty is that floorplans are compact and minimalist, so repetitive hallways, corners, and room outlines create severe pose ambiguity (Chen et al., 2024, Grader et al., 12 Jul 2025, Chen et al., 2 Aug 2025, Wüest et al., 14 Sep 2025).

1. Problem formulation and map representations

A common formulation models FLoc as probabilistic inference over a discretized pose space. Given an observation $O_{I,F}=(I,F)$ or $O_{\mathcal{I},F}=(\mathcal{I},F)$, the objective is to estimate

$\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$

where $\mathcal S$ is a finite set of candidate poses obtained by discretizing continuous $SE(2)$. This formulation appears explicitly in ray-based systems such as F$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$0Loc and in later extensions that preserve the same posterior-search structure while changing the observation model (Chen et al., 2024, Grader et al., 12 Jul 2025, Chen et al., 2 Aug 2025).

The floorplan itself is represented in several different ways. In F$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$1Loc and related geometry-driven systems, the map is primarily an occupancy representation containing walls and doors, and the observation model compares predicted geometric structure against floorplan-derived reference structure (Chen et al., 2024). Later work makes the floorplan semantic, treating it as $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$2 with classes such as wall, window, and door, or with room polygons that can be used as masks (Grader et al., 12 Jul 2025). Other methods move away from exact vector geometry: SceneAligner explicitly targets rasterized or symbolic floorplans by aligning them to a reconstruction-derived density-map proxy, while Z-FLoc matches geometric primitives extracted from a BEV reconstruction to floorplan primitives through a 2D similarity transform (Cho et al., 21 May 2026, Umemura et al., 3 Jun 2026).

Not all formulations search directly over a discretized $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$3 grid. LaLaLoc reduces the target to 2 DoF $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$4, interprets the floorplan as an extruded 3D layout model, and localizes panoramas through a shared latent layout space and direct latent pose optimization (Howard-Jenkins et al., 2021). SceneAligner and Z-FLoc instead formulate localization as cross-modal alignment under $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$5, reflecting the fact that monocular or foundation-model reconstructions may have imperfect metric scale (Cho et al., 21 May 2026, Umemura et al., 3 Jun 2026). This suggests that contemporary FLoc is less a single algorithmic template than a family of cross-modal pose-estimation problems linked by a common prior map.

2. Core observation models: latent layouts, latent rendering, and ray scans

Early modern FLoc systems differ primarily in how they represent the visual observation so that it becomes comparable to a floorplan. LaLaLoc constructs a shared latent embedding space for RGB panoramas and layouts rendered from a floorplan-derived 3D model. Coarse localization is performed by nearest-neighbor retrieval in that latent space, after which pose refinement proceeds either by local Vogel Disc re-sampling or by differentiable latent pose optimization,

$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$6

where $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$7 is the panorama embedding and $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$8 is the renderer. On Structured3D, LaLaLoc reports a median localization error of $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$9 cm in the RGB/furnished setting, with pose recall@1 of $(x,y)$0 and within-1 m accuracy of $(x,y)$1 (Howard-Jenkins et al., 2021).

LASER preserves the Monte Carlo Localization viewpoint but replaces explicit per-pose image synthesis with latent space rendering. A floor map is rasterized into boundary points, each endowed with geometry-conditioned codebooks, and a pose hypothesis is rendered directly into a circular feature $(x,y)$2 whose indexing is rotation-aware. The shared metric space accommodates both map hypotheses and panorama or perspective image queries, and the paper reports sampling speed above $(x,y)$3 KHz together with strong perspective and panorama performance on ZInD and Structured3D (Min et al., 2022). In LASER, orientation is naturally represented as a cyclic shift of the circular descriptor, so yaw search becomes a structured correlation problem rather than a generic embedding comparison.

F$(x,y)$4Loc establishes the ray-based probabilistic paradigm that many later papers extend. Its observation module predicts floorplan depth rather than pose directly: gravity-aligned RGB is processed by single-view and multi-view branches, their depth distributions are fused by a learned selector,

$(x,y)$5

and the result is converted into an equiangular 1D ray scan. The floorplan is pre-rendered into corresponding ray scans, and the observation likelihood becomes

$(x,y)$6

A histogram filter over $(x,y)$7 then fuses these observations over time, implemented efficiently as grouped convolution. F$(x,y)$8Loc reports $(x,y)$9 Hz histogram-filter iterations, handles non-upright cameras by virtual roll-pitch augmentation, and improves substantially over earlier floorplan baselines on Gibson and Structured3D (Chen et al., 2024).

These observation models reflect three recurrent design choices in FLoc: embedding-based cross-modal retrieval, structured latent rendering, and explicit geometric ray matching. Later work largely retains one of these choices and addresses its failure modes rather than discarding it outright.

3. Disambiguation through semantics, context, and contrastive priors

A central claim in recent FLoc research is that geometry alone is often insufficient. The semantic-ray framework of “Supercharging Floorplan Localization with Semantic Rays” keeps the F$\theta$0Loc-style probabilistic pose volume but augments depth rays $\theta$1 with semantic rays $\theta$2. It constructs a depth probability volume $\theta$3, a semantic probability volume $\theta$4, and fuses them as

$\theta$5

Because semantic labels are discrete, the method replaces linear interpolation with majority-vote interpolation, performs coarse-to-fine refinement only on Top-$\theta$6 candidates, and can optionally mask the posterior with a room polygon when room-type confidence exceeds a threshold. On S3D it reports $\theta$7 and $\theta$8 for the room-aware variant, versus $\theta$9 and $\phi$0 for F$\phi$1Loc; on ZInD it reports $\phi$2 and $\phi$3, versus $\phi$4 and $\phi$5 for F$\phi$6Loc (Grader et al., 12 Jul 2025). The paper explicitly notes that hard ray assignment and hard room masking outperform soft alternatives.

A related but label-free line of work injects scene context into FLoc through pretraining. “Perspective from a Broader Context” argues that room style provides a prior over plausible floorplan regions. It learns a room discriminator from unlabeled room images using a constraint matrix derived from scene and episode metadata, InfoMap clustering, a cluster-level contrastive loss, and a binary style-pair prediction loss. The pretrained encoder is then transferred into the F$\phi$7Loc observation model, yielding consistent gains over F$\phi$8Loc and 3DP on Gibson and Structured3D(full) (Chen et al., 2 Aug 2025). “Perspective from a Higher Dimension” instead uses self-supervised 3D geometric priors: Geometry-Constrained View Invariance (GCVI) and View-Scene Aligned Geometric (VSAG) priors are learned on ScanNet and then injected into the F$\phi$9Loc encoder without increasing inference-time burden. On Structured3D, the resulting single-frame model improves $O_{I,F}=(I,F)$0 from $O_{I,F}=(I,F)$1 to $O_{I,F}=(I,F)$2 and $O_{I,F}=(I,F)$3 from $O_{I,F}=(I,F)$4 to $O_{I,F}=(I,F)$5 (Chen et al., 25 Jul 2025).

DisCo-FLoc makes the ambiguity problem explicit by separating candidate generation from disambiguation. A depth-aware ray-regression predictor first produces a Depth-Aware FLoc Probabilistic Map; a second stage then uses dual-level visual-geometric contrastive learning with position-level and orientation-level negatives to re-rank the top candidates. The paper argues that ambiguity is a matching problem under structural repetition rather than merely a depth-estimation problem. On Structured3D(full), it reports $O_{I,F}=(I,F)$6 at $O_{I,F}=(I,F)$7 m, $O_{I,F}=(I,F)$8 m, $O_{I,F}=(I,F)$9 m, and $O_{\mathcal{I},F}=(\mathcal{I},F)$0 m+$O_{\mathcal{I},F}=(\mathcal{I},F)$1, and it states that it outperforms semantic-based variants without requiring semantic labels (Meng et al., 5 Jan 2026). A plausible implication is that strict pose-conditioned contrastive objectives can substitute for hand-annotated semantics when repeated geometry is the dominant error source.

The proof-of-concept COMPASS points in a similar direction but at descriptor level. It encodes the floorplan into a $O_{\mathcal{I},F}=(\mathcal{I},F)$2 radial descriptor with normalized range, hit type, range gradient, inverse range, and local range variance, and constructs an image-side hit-type descriptor from dual fisheye images by detecting windows and projecting them to azimuth bins. On a known-pose sample from the Hilti-Trimble SLAM Challenge 2026 dataset, cross-correlation of the hit-type channels peaks at $O_{\mathcal{I},F}=(\mathcal{I},F)$3 with a shift of $O_{\mathcal{I},F}=(\mathcal{I},F)$4, validating wall-window pattern matching as a structural cue, though not yet a full localization system (Shaheer et al., 28 Apr 2026).

4. Uncertainty, reconstruction-grounded alignment, and zero-shot geometry

Another major development is uncertainty-aware FLoc. UnLoc keeps the FLoc/F$O_{\mathcal{I},F}=(\mathcal{I},F)$5Loc sequence-localization structure—predict depth, compare against floorplan rays, fuse over time—but replaces deterministic depth with a Laplace-distributed model. For each image column it predicts depth $O_{\mathcal{I},F}=(\mathcal{I},F)$6 and uncertainty $O_{\mathcal{I},F}=(\mathcal{I},F)$7, trains with

$O_{\mathcal{I},F}=(\mathcal{I},F)$8

and constructs the observation likelihood as a product of Laplace factors. The method uses off-the-shelf pretrained monocular depth encoders, with Depth Anything V2 highlighted as the preferred backbone, rather than environment-specific depth training. On LaMAR HGE, the reported success rate rises from $O_{\mathcal{I},F}=(\mathcal{I},F)$9 to $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$0 for 15-frame sequences and from $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$1 to $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$2 for 100-frame sequences; on LaMAR CAB, UnLoc generalizes substantially better than F$\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$3Loc (Wüest et al., 14 Sep 2025). The key significance is not only better accuracy but explicit propagation of observation confidence into the Bayesian posterior.

Several recent methods depart from exhaustive pose search and instead reconstruct a scene-derived proxy that is subsequently aligned to the floorplan. SceneAligner reconstructs a gravity-aligned 3D scene from an unconstrained image collection, filters the point cloud to retain reliable vertical structures, projects it to a top-down density map $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$4, and estimates a similarity transform $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$5 between that proxy and the floorplan. To bridge the appearance gap between density maps and floorplans, it adapts DINOv3 ViT-B/16 with LoRA and a loss that combines feature matching, coordinate regression, topology preservation, and geometry consistency. On the in-the-wild C3 dataset, it reports combined angular-positional recall $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$6 of $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$7, compared with $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$8 for C3Po and $\hat S=\arg\max_{S_i\in\mathcal S} p(S_i\mid O),$9 for plain DINOv3; on Structured3D it reports combined recall $\mathcal S$0, compared with $\mathcal S$1 for F$\mathcal S$2Loc and $\mathcal S$3 for UnLoc (Cho et al., 21 May 2026).

Z-FLoc is a zero-shot alternative to learning-based FLoc. It reconstructs a BEV map from monocular images, extracts lines and circles, and matches them to floorplan primitives using minimal solvers embedded in a hybrid RANSAC framework. The alignment is a 2D similarity transform,

$\mathcal S$4

scored by a consistency term $\mathcal S$5 and a violation term $\mathcal S$6 that penalizes floorplan walls placed inside observed free space. On LaMAR CAB, a strict cross-scene setting, Z-FLoc reports $\mathcal S$7 SR@1m at $\mathcal S$8 and $\mathcal S$9 frames, whereas F$SE(2)$0Loc-mono reports $SE(2)$1 across all lengths and UnLoc reports $SE(2)$2 at $SE(2)$3 frames and $SE(2)$4 at $SE(2)$5 frames (Umemura et al., 3 Jun 2026). This suggests that geometry-only floorplan alignment can be more robust than learned cross-modal matching under strong domain shift, albeit with dependence on reconstruction quality and structural primitive availability.

PALMS$SE(2)$6 occupies a distinct modular corner of the design space. It reconstructs scale-corrected 3D geometry from posed RGB images using Depth Pro, projects points near camera height into a 2D layout representation, extracts line segments, and matches them to the floorplan by convolution with a recorded-wall kernel and a Certainly Empty Space penalty: $SE(2)$7 It requires no training, outputs a three-dimensional pose posterior, and can initialize a particle filter for sequential localization. On a custom campus dataset, PALMS$SE(2)$8 reports $SE(2)$9 full-view localization accuracy versus $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$00 for PALMS and $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$01 for F3Loc, and on 33 trajectories it reduces RMSE from $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$02 m for F3Loc and $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$03 m for PALMS to $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$04 m (Cheng et al., 12 Nov 2025).

5. Benchmarks, metrics, and empirical trends

The benchmark ecosystem for FLoc is heterogeneous but increasingly standardized around a small set of datasets. Structured3D and Gibson variants dominate synthetic evaluation; ZInD provides real residential panoramas and perspective crops; LaMAR HGE and CAB stress real-world sequence localization and cross-building generalization; C3 targets in-the-wild floorplan alignment; and some methods introduce campus-scale or challenge-specific datasets for scan-based or fisheye settings (Chen et al., 2024, Grader et al., 12 Jul 2025, Wüest et al., 14 Sep 2025, Cho et al., 21 May 2026, Cheng et al., 12 Nov 2025, Shaheer et al., 28 Apr 2026).

Metrics also reflect methodological differences. Ray-based and sequence-based systems commonly report recall at $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$05 m, $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$06 m, $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$07 m, and $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$08 m with $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$09 orientation tolerance; Gibson(t) and LaMAR-style tracking settings additionally report success rate over the final 10 frames and RMSE; SceneAligner reports angular recall, positional recall as a percentage of floorplan diagonal, combined angular-positional recall, PCK, and RMSE; stationary scan methods use heatmap-based direct localization accuracy and trajectory-level error (Chen et al., 2024, Chen et al., 2 Aug 2025, Wüest et al., 14 Sep 2025, Cho et al., 21 May 2026, Cheng et al., 12 Nov 2025).

Across the reported settings, several empirical trends are consistent. First, geometry-only floorplan localization remains competitive when the geometry is represented effectively: F$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$10Loc improves over LASER on Structured3D monocular evaluation, and Z-FLoc exceeds strong learned baselines on unseen buildings (Chen et al., 2024, Umemura et al., 3 Jun 2026). Second, disambiguation cues produce the largest jumps when layouts are repetitive: semantic rays, room-style priors, 3D geometric priors, and dual-level contrastive disambiguation all report substantial gains over geometry-only baselines (Grader et al., 12 Jul 2025, Chen et al., 2 Aug 2025, Chen et al., 25 Jul 2025, Meng et al., 5 Jan 2026). Third, uncertainty matters most on hard real scenes rather than easier synthetic ones, as emphasized by UnLoc’s much larger gains on LaMAR HGE than on Gibson(t) (Wüest et al., 14 Sep 2025). Fourth, the field is broadening from precise vector floorplans and bounded indoor apartments toward raster floorplans, large public buildings, campus environments, and unconstrained photo collections (Cho et al., 21 May 2026, Cheng et al., 12 Nov 2025).

Direct numerical comparisons must be interpreted with care because protocols differ. Even so, the aggregate direction is clear in the reported results: on Structured3D-style benchmarks, F$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$11Loc reports $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$12 combined recall at $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$13 m+$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$14, SceneAligner reports $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$15, and DisCo-FLoc reports $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$16 on Structured3D(full); on S3D, the semantic-ray method reports $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$17 in the same $\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$18 m+$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$19 recall metric (Chen et al., 2024, Cho et al., 21 May 2026, Meng et al., 5 Jan 2026, Grader et al., 12 Jul 2025). This suggests steady progress, but the variation in inputs—single image, sequence, stationary scan, or unconstrained collection—remains a major axis of non-comparability.

6. Scope, adjacent problems, and recurring limitations

The term FLoc is sometimes used loosely, and several neighboring problems are better understood as floorplan-conditioned localization or refinement rather than direct visual floorplan localization. “Vision-Based Localization and LLM-based Navigation for Indoor Environments” performs waypoint classification from smartphone imagery and uses a cleaned floorplan image only for LLM-based route generation; it is explicitly described as not a pure floorplan-to-location retrieval method in the usual FLoc sense (Rahimi et al., 11 Aug 2025). “Floorplan-Aware Camera Poses Refinement” uses a technical floorplan as a structural prior in RGB-D bundle adjustment to refine already-estimated poses rather than localize from scratch (Sokolova et al., 2022). Fusion-DHL refines IMU-WiFi trajectories with a floorplan-conditioned CNN, and FP-Loc is a LiDAR-specific floor-plan localization system whose conceptual relation to FLoc is strong but whose sensing and optimization pipeline are not vision-based (Herath et al., 2021, Gao et al., 2022). A common misconception is therefore to equate any floorplan-aware localization or navigation system with visual FLoc proper.

The limitations reported across the literature are also recurrent. Repetitive geometry remains the canonical failure mode. Learning-based methods are sensitive to domain shift, environment-specific training, or depth-prediction quality; uncertainty-aware methods alleviate but do not eliminate this dependence (Wüest et al., 14 Sep 2025). Semantic methods benefit from doors, windows, and room types, but some papers explicitly note that semantic annotations are expensive and limited in availability, motivating contrastive or self-supervised substitutes (Grader et al., 12 Jul 2025, Meng et al., 5 Jan 2026). Reconstruction-grounded methods depend on gravity estimation and 3D reconstruction quality, and SceneAligner notes that errors in geometry propagate into the density map and then into alignment; Z-FLoc similarly depends on reliable BEV structure and observable geometric primitives (Cho et al., 21 May 2026, Umemura et al., 3 Jun 2026). PALMS$\mathbf{s}_t=[s_{x,t},s_{y,t},s_{\phi,t}]$20 identifies transparent surfaces and residual scale errors as important failure sources for monocular-depth-based scan matching (Cheng et al., 12 Nov 2025).

A broader pattern is that the field is moving from direct cross-modal retrieval toward richer structural mediation. That mediation may take the form of rays, uncertainty distributions, semantic channels, learned context priors, geometric primitives, or 3D-grounded density maps. The shared objective remains unchanged: to bridge the domain gap between appearance-heavy camera observations and sparse architectural plans without sacrificing the compactness and long-term stability that make floorplans attractive priors in the first place.