2D Floorplan Localization (FLoc)

• 2D Floorplan Localization is the task of estimating camera or robot poses within a known architectural floorplan using diverse sensor inputs.
• It integrates geometric registration, learned cross-modal representations, and probabilistic filtering to bridge the gap between top-down maps and egocentric observations.
• The approach addresses challenges like repeated structures and occlusions by leveraging semantic cues and uncertainty modeling for reliable localization.

Searching arXiv for recent and foundational papers on 2D floorplan localization to ground the article. Search query: floorplan localization visual localization floorplan camera floorplan arXiv 2D floorplan localization (FLoc) is the problem of estimating a camera or robot pose in the coordinate frame of a known architectural floorplan from observations such as RGB images, panoramas, dual fisheye imagery, monocular sequences, LiDAR scans, or reconstructed 3D geometry. Its appeal follows from the fact that floorplans are readily available, compact, long-term persistent, and robust to changes in visual appearance, while its central difficulty is the severe domain gap between egocentric sensing and a sparse top-down map. That gap is amplified by repeated room geometries, partial observability, and discrepancies between as-designed and as-built structure (Chen et al., 2024, Umemura et al., 3 Jun 2026).

1. Problem formulation and pose spaces

FLoc is usually posed in the plane, but the exact state depends on sensor model and inference strategy. LaLaLoc performs camera relocalization with respect to a known 2D floorplan and predicts a 2-DoF camera pose on the floorplan, with orientation fixed or known for rendering (Howard-Jenkins et al., 2021). F$^3$Loc explicitly estimates the SE(2) state $\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$ and maintains a posterior over a discretized grid of translations and orientations (Chen et al., 2024). COMPASS targets planar position $\mathbf{p}=(t_x,t_y)$ and yaw $\psi$ in a 2D floorplan coordinate frame, while also recovering roll and pitch through vanishing-point estimation (Shaheer et al., 28 Apr 2026). FP-Loc uses a 2D CAD floor plan as prior but outputs full 6-DoF LiDAR localization by solving roll, pitch, and height from ceiling and ground planes and optimizing planar registration against vertical structure elements (Gao et al., 2022).

The input assumptions vary accordingly. Some systems consume a single RGB panorama and a known floorplan, as in LaLaLoc (Howard-Jenkins et al., 2021). Others operate on perspective RGB images plus gravity and ego-motion, as in F$^3$Loc (Chen et al., 2024), or on a monocular image sequence with known relative SE(3) poses, as in Z-FLoc (Umemura et al., 3 Jun 2026). SceneAligner takes an unconstrained image collection and a raster floorplan, reconstructs a gravity-aligned 3D scene, and estimates a 2D similarity transform that localizes all reconstructed cameras in floorplan coordinates (Cho et al., 21 May 2026). The common denominator is that localization is tied to a pre-existing 2D plan rather than to a pre-built image database or a dense 3D map.

This variety has produced several adjacent formulations. In some work, FLoc is a single-shot retrieval-and-refinement problem. In others, it is a sequential Bayesian filtering problem over SE(2). In reconstruction-oriented settings such as SceneAligner or SALVe, the primary estimate is a scene-to-floorplan alignment or a set of globally consistent panorama poses, from which floorplan-localized camera poses follow directly (Cho et al., 21 May 2026, Lambert et al., 2024).

2. Map, scene, and observation representations

The defining design choice in FLoc is how the floorplan prior is rendered into a signal that can be compared to a camera observation. F$^3$Loc uses a ray-based representation: for each hypothesized pose it casts equiangular rays into the occupancy map and compares them to rays predicted from RGB images (Chen et al., 2024). LaLaLoc extrudes the 2D floorplan into a simple 3D model of walls, floors, and ceilings, renders layout depth panoramas, and embeds both rendered layouts and RGB panoramas into a shared 128-D latent layout space (Howard-Jenkins et al., 2021). COMPASS instead defines a multi-channel radial descriptor $\mathbf{D}\in\mathbb{R}^{5\times 360}$ whose channels are normalized range, hit type, range gradient, inverse range, and local range variance, with hit type explicitly distinguishing wall, window, and open (Shaheer et al., 28 Apr 2026). Z-FLoc reduces both map and observation to geometric primitives—lines and circles—extracted from a floorplan and from a BEV wall map reconstructed from monocular imagery (Umemura et al., 3 Jun 2026). SceneAligner projects a gravity-aligned 3D reconstruction into a 2D density map and learns correspondences between this proxy and a raster floorplan (Cho et al., 21 May 2026). “Semantic rays” extend the ray paradigm by predicting not only depth but also wall, door, and window labels, and optionally room labels, from a single image (Grader et al., 12 Jul 2025).

System	Floorplan-side signal	Observation-side signal
LaLaLoc (Howard-Jenkins et al., 2021)	Extruded 3D floorplan rendered as layout depth	RGB panorama embedded in latent layout space
F$^3$Loc (Chen et al., 2024)	Occupancy-map ray casts	Single-view or multi-view floorplan depth rays
COMPASS (Shaheer et al., 28 Apr 2026)	$5\times 360$ semantic-structural radial descriptor	Dual-fisheye descriptor, currently validated on hit type
Z-FLoc (Umemura et al., 3 Jun 2026)	2D lines and circles from floorplan	BEV wall primitives from monocular reconstruction
SceneAligner (Cho et al., 21 May 2026)	Raster floorplan	Gravity-aligned density map from reconstructed 3D scene
Semantic Rays (Grader et al., 12 Jul 2025)	Semantic floorplan with wall, door, window, room labels	Depth rays and semantic rays from a single image

These representations expose different invariances. Radial descriptors and ray scans make rotation an angular shift, which simplifies heading search. Latent layout spaces enable cross-modal retrieval and differentiable pose refinement. Primitive and density-map formulations reduce the visual-floorplan gap by converting perspective imagery into top-down geometric proxies before alignment. Semantic encodings exploit the fact that floorplans may carry information beyond occupancy alone, including windows, doors, openings, room labels, or textual cues (Shaheer et al., 28 Apr 2026, Zimmerman et al., 2024).

3. Dominant algorithmic paradigms

One major family is direct geometric registration. FP-Loc uses robust ceiling and ground plane detection to solve roll, pitch, and height, extracts vertical wall and pillar features, queries floor-plan elements through an Approximate Nearest Neighbour Field, and optimizes a pair-wise regularized windowed pose graph (Gao et al., 2022). Z-FLoc follows a different geometric route: it extracts lines and circles from a BEV reconstruction and from the floorplan, estimates a 2D similarity transform with minimal solvers such as 3L, 2L, 2C, and LC, and scores hypotheses with both alignment consistency and free-space violation terms inside a hybrid RANSAC framework (Umemura et al., 3 Jun 2026).

A second family uses learned cross-modal representations. LaLaLoc first performs coarse retrieval over rendered candidate poses, then refines with Vogel Disc Re-sampling and latent pose optimisation by minimizing the latent distance between a rendered layout and an RGB panorama through a differentiable renderer (Howard-Jenkins et al., 2021). LASER replaces explicit rendering with latent space rendering: it learns geometry-conditioned codebooks attached to floor-map points, renders pose hypotheses directly into a circular latent feature, treats heading as a circular shift, and uses an MCL-style measurement model built from cosine similarity (Min et al., 2022).

A third family casts FLoc as probabilistic filtering over ray-based observation models. F$^3$Loc predicts floorplan depth from a single image and from a short multi-view window, fuses the resulting depth distributions, converts them into equiangular rays, and evaluates an observation likelihood given by the exponential of the $\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$0 distance between predicted and map rays (Chen et al., 2024). UnLoc retains the same floorplan-ray logic but models each predicted depth ray as a Laplace distribution with learned scale, making the observation model explicitly uncertainty-aware and integrating it with a histogram filter over SE(2) (Wüest et al., 14 Sep 2025). PALMS+ uses a different observation generator—scale-aligned 3D point clouds from posed RGB images and a foundation monocular depth model—but the back end is again a pose posterior over discretized $\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$1, obtained by convolving a rasterized line-and-CES kernel over the floorplan and optionally propagated through a particle filter (Cheng et al., 12 Nov 2025).

A related line uses pairwise alignment and graph optimization to localize sparse panoramic observations in a common 2D frame. SALVe detects windows, doors, and openings in panoramas, proposes pairwise relative poses from these semantics, verifies each candidate with a learned BEV alignment classifier, and then optimizes a pose graph in GTSAM before stitching room layouts into a floorplan (Lambert et al., 2024). Although reconstruction is its stated goal, the intermediate problem is explicitly the localization of panoramas in a common 2D floorplan-like frame.

4. Empirical profile of representative systems

Reported results are benchmark-specific and sensor-specific, so they should not be collapsed into a single leaderboard. They nonetheless show the range of what modern FLoc systems can achieve, from centimeter-level localization in synthetic or domestic settings to strong zero-shot performance on cross-building sequential benchmarks, and to proof-of-concept semantic-structural alignment from floorplan windows alone.

System	Representative result	Setting
LaLaLoc (Howard-Jenkins et al., 2021)	Localise a single RGB panorama image to within 8.3cm	Domestic environment, only floor plan as prior
LASER (Min et al., 2022)	ZInD panoramas: med terr 5.16 cm; 1m recall 97.12%	Panorama queries on metric floor maps
F$\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$2Loc (Chen et al., 2024)	Gibson(t): success rate @ 1 m 94.6%; RMSE 0.12 m on successful runs	Fusion plus histogram filtering
Semantic Rays (Grader et al., 12 Jul 2025)	R@1m30° 57.49 on Structured3D; 31.86 on ZInD	Single-image semantic-aware localization
UnLoc (Wüest et al., 14 Sep 2025)	LaMAR HGE Original, T=100: SR@1m 100%; CAB transfer, T=100: SR@1m 50.0%	Sequential uncertainty-aware localization
PALMS+ (Cheng et al., 12 Nov 2025)	Custom campus full-view Acc.@1m30° 30.4%, or 38.0% with masked depth; final localization error 1.3 m for PALMS+*	Stationary scan and sequential particle filtering
SceneAligner (Cho et al., 21 May 2026)	Combined recall @ (30°, 20%) 73.58 on C3; 51.6 @ (30°, 1m) on Structured3D	3D-grounded alignment to raster floorplans
Z-FLoc (Umemura et al., 3 Jun 2026)	Gibson(t): SR@1m 99.5% @100 frames; LaMAR CAB: 100% @100 and 50 frames	Zero-shot primitive-based localization
COMPASS (Shaheer et al., 28 Apr 2026)	Correlation peaks at 0.9486; 231 out of 360 bins agree; peak at 0° shift	Single-pose proof-of-concept heading alignment

The numerical spread reflects differing regimes rather than only differing algorithmic strength. LaLaLoc and LASER assume panorama-centric inputs; PALMS+ is built around a stationary rotational scan; SceneAligner estimates a scene-level 2D alignment from sparse or unconstrained image collections; COMPASS currently validates cross-modal matching at a single known pose and only on the hit-type channel. Even so, the aggregate record shows several consistent patterns: semantic augmentation improves recall in ray-based systems, uncertainty modeling improves convergence on short and long sequences, and zero-shot geometric pipelines can remain competitive on unseen buildings.

5. Ambiguity, robustness, and recurrent failure modes

The core ambiguity in FLoc is structural repetition. Rectangular rooms, corridor segments, and parallel walls often produce similar occupancy-only signatures. LaLaLoc explicitly reports failure cases in which the wrong room is retrieved but still yields a plausible alignment, and the semantic-ray formulation is motivated by repeated layouts that remain ambiguous under depth-only matching (Howard-Jenkins et al., 2021, Grader et al., 12 Jul 2025). This is why several systems move beyond occupancy to wall-window-opening patterns, room labels, or textual cues.

A common misconception is that floorplans contribute only geometry. Multiple systems contradict that assumption directly. COMPASS treats the floor plan as two binary raster masks, $\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$3 for walls and $\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$4 for glazing, and uses the resulting wall–window–opening pattern as part of the descriptor (Shaheer et al., 28 Apr 2026). The semantic-ray framework expects wall, door, and window labels in the floorplan and can further mask the probability volume with room labels (Grader et al., 12 Jul 2025). Long-term localization work based on abstract semantic maps overlays floor plans with room segmentation, semantic rectangles, and text likelihood maps, and reports that integrating textual cues improves localization stability when localizing in environments with high geometric symmetry and lack of semantic features (Zimmerman et al., 2024).

Robustness is constrained by the fidelity of the prior and by the visibility of the relevant structure. COMPASS explicitly notes plan-versus-as-built mismatch in a construction site, sensitivity to occluded windows, and the limited discriminative power of the hit-type channel in interior corridors without windows (Shaheer et al., 28 Apr 2026). PALMS+ shows that scale alignment and masking depth artifacts behind glass materially affect localization, and its success rates fall in repetitive hallways even when long-range geometry is available (Cheng et al., 12 Nov 2025). SceneAligner states that failure or poor quality in 3D reconstruction directly degrades the density map and thus the floorplan alignment (Cho et al., 21 May 2026). UnLoc addresses an adjacent issue by treating monocular depth as a distribution rather than a deterministic quantity, flattening the contribution of rays that pass through glass, textureless walls, or otherwise unreliable regions (Wüest et al., 14 Sep 2025).

The principal methodological controversy is the trade-off between learned cross-modal models and zero-shot geometric pipelines. Z-FLoc positions large-scale training data and environment-specific retraining as a practical limitation of existing learning-based approaches and reports strong cross-building gains from geometry-only matching (Umemura et al., 3 Jun 2026). Learned systems, however, often exploit complementary priors—latent layout similarity, semantic rays, uncertainty estimation, or broader scene context—that pure geometry does not encode. The current literature therefore does not support a single universal winner; instead, it exposes a division between regimes where trainable structural priors dominate and regimes where analytic geometry generalizes more cleanly.

6. Emerging directions

One current direction is the explicit exploitation of structural semantics already present in floorplans. The semantic-ray framework jointly estimates depth and semantic rays and builds a structural-semantic probability volume, while COMPASS designs a compact rotation-equivariant descriptor whose hit-type channel encodes wall, window, and opening structure around the pose (Grader et al., 12 Jul 2025, Shaheer et al., 28 Apr 2026). This suggests a shift from occupancy-only matching toward descriptors that preserve both geometry and architectural category.

A second direction is the import of priors that are not themselves floorplan-native. “Perspective from a Higher Dimension” injects 3D geometric priors into the visual encoder through multi-view and view-scene aligned self-supervision, then plugs that encoder into an unchanged F$\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$5Loc pipeline (Chen et al., 25 Jul 2025). “Perspective from a Broader Context” uses unsupervised room-style clustering to pre-train a room discriminator and then transfers that encoder into FLoc, arguing that broader scene context can eliminate ambiguities that survive 2D and 3D geometric reasoning (Chen et al., 2 Aug 2025). These works do not change the formal output of FLoc, but they alter the visual statistics that the observation model can exploit.

A third direction is in-the-wild and zero-shot deployment. SceneAligner reformulates localization as 3D-grounded 2D alignment between a reconstructed density map and an arbitrary raster floorplan, dispensing with discretized pose search and vectorized-plan assumptions (Cho et al., 21 May 2026). Z-FLoc argues that lines and circles in a BEV reconstruction provide appearance-invariant structural constraints that generalize without retraining across novel environments (Umemura et al., 3 Jun 2026). SALVe, from the reconstruction side, shows that semantic alignment verification over sparse panoramas can localize cameras in a common 2D frame before floorplan stitching (Lambert et al., 2024). A plausible implication is that future FLoc systems will increasingly operate on intermediate top-down proxies—density maps, BEV reconstructions, primitive graphs, or semantic radial signatures—rather than on raw perspective images alone.

A fourth direction concerns temporal and probabilistic integration. F$\mathbf{s}_t = [s_{x,t}, s_{y,t}, s_{\phi,t}]$6Loc, UnLoc, PALMS+, and long-term semantic MCL systems all use explicit Bayesian filtering or particle filtering, and COMPASS explicitly lists temporal integration and probabilistic frameworks as future work for semantic-structural matching (Chen et al., 2024, Wüest et al., 14 Sep 2025, Cheng et al., 12 Nov 2025, Shaheer et al., 28 Apr 2026). This suggests that the frontier is not only better single-frame matching, but also better handling of uncertainty, multi-modality, and map discrepancy over trajectories and over long time spans.

Taken together, recent work frames FLoc less as a single algorithmic recipe than as a family of cross-modal alignment problems. The floorplan may be queried as an occupancy field, a latent layout renderer, a semantic ray source, a set of geometric primitives, a density-map target, or an abstract semantic map with text. The observation may be a single perspective view, a panorama, a stationary scan, a fisheye pair, a monocular sequence, or an unconstrained image collection. The continuing technical theme is the same: reducing the mismatch between what a floorplan encodes and what a camera or LiDAR actually sees.