What if a robot could localize indoors using only a floorplan, without building a giant image or ray database first? FreeLoc makes that idea practical by querying geometry online and refining rough guesses with diffusion. Floorplans are compact and widely available maps for indoor localization, but existing high-performing methods still turn them into dense scene-specific offline databases.

That database ties accuracy, storage, and runtime to the sampling resolution of the discretized pose space. FreeLoc instead treats the floorplan as a directly queryable geometric map for online RGB-based localization. FreeLoc queries floorplan rays on the fly to retrieve plausible pose anchors, then uses diffusion-aided refinement to produce accurate continuous pose estimates.

For sequential localization, it constructs online likelihoods from coarse-sampled candidates and refined pose hypotheses, enabling histogram-filter-based temporal fusion without offline databases. Autonomous exploration. In new indoor scenes, localization should be accurate and readily deployable with minimal scene-specific preparation.

Traditional methods typically rely on pre-built image databases or accurate three-dimensional scene models. Those representations are costly to construct, storage-intensive, and difficult to maintain, which makes scaling across buildings and changing environments difficult.

Floorplans offer a lightweight and scalable representation: compared with image databases and three-dimensional scene models, they are compact, widely available, and relatively stable over time. Ideally, a floorplan would be a directly queryable geometric map, allowing localization without an additional scene-specific database.

Existing high-performing methods still discretize the SE two pose space, precompute dense pose-indexed observations, and match query images against those entries. Because localization precision, storage, and runtime scale with sampling resolution, the discretized paradigm creates a trade-off among accuracy, efficiency, and scalability.

The central question is whether a system given only a floorplan can directly query it online for accurate RGB-based single-frame and sequential localization in new environments, without an offline database. The paper draws a coarse-to-fine intuition from human localization: narrow the search to a few plausible regions, then refine the location by matching local geometric details with the current observation.

Making that intuition work online is non-trivial because the system must rapidly search the entire floorplan pose space for plausible initial hypotheses. Retrieved poses are approximate because finite sampling creates discretization errors, while noisy and incomplete RGB-derived observations limit matching accuracy.

Sequential localization also needs observation likelihoods over the discretized pose space, but those likelihoods can no longer be densely evaluated from a precomputed scene-specific ray database. FreeLoc treats the floorplan as a directly queryable geometric map rather than a scene-specific offline retrieval database.

It first retrieves a compact set of plausible pose anchors by querying geometric rays from the floorplan on the fly with a GPU-accelerated rotation-aware ray kernel. A diffusion-aided pose refinement network estimates residual pose corrections from the discrepancy between image-side and floorplan-side ray observations.

Instead of one-shot deterministic regression, pose refinement is an iterative denoising process that progressively corrects imperfect coarse anchors under ray-level geometric guidance. This refinement moves beyond the resolution of coarse online samples and produces accurate pose hypotheses in continuous pose space without exhaustive dense sampling or precomputed scene-specific ray databases.

For sequential localization, FreeLoc integrates the online coarse-to-fine estimator into a probabilistic filtering pipeline. It constructs observation likelihoods online from coarse-sampled candidates and refines their high-confidence regions using refined pose hypotheses before Bayesian filtering.

That online construction bridges single-frame localization and temporal fusion without reintroducing a scene-specific database. The top row shows a sequence of RGB observations paired with a floorplan, while the lower panels visualize the resulting poses. Across the single-frame examples, the green ground-truth and blue estimated arrows remain closely aligned, and the sequential panel traces a continuous path that follows the reference trajectory through the floorplan.

The task is RGB-based camera pose estimation in a known two-dimensional floorplan. FreeLoc treats the floorplan as an online geometric query interface rather than a precomputed pose-indexed database. Given a floorplan, camera intrinsics, gravity direction, and either one RGB image or an RGB sequence with inter-frame relative poses, the goal is to estimate camera pose.

The estimated pose is in the floorplan coordinate system: two-dimensional camera position plus yaw angle. The unified framework retrieves pose anchors through on-the-fly floorplan ray querying, refines them with diffusion-aided residual optimization, and reuses the hypotheses to construct online observation likelihoods.

This supports both single-frame and sequential localization without scene-specific offline ray databases. The top row shows the coarse-to-fine handoff: sampled floorplan poses produce a sparse set of ray matches, which then passes through RDE and diffusion-guided refinement before reaching the final pose.

In the bottom row, the sparse likelihood is visibly smoothed into a broader spatial field, then updated through Bayesian filtering as ego motion incorporates temporal information. The first stage retrieves a compact set of plausible pose anchors by querying floorplan geometry online instead of matching the image against a precomputed pose-indexed ray database.

This provides broad pose-space coverage for ambiguous indoor layouts while keeping the candidate set sparse enough for efficient refinement. Because practical RGB images can come from handheld or head-mounted devices with arbitrary roll and pitch, each image is first normalized into a gravity-aligned view.

Roll and pitch can come from inertial sensors or be estimated with single-image calibration methods such as GeoCalib when sensor readings are unavailable. The image is rectified with a homography that aligns its vertical direction with gravity around the optical center.

The rectified image and a visibility mask are used to estimate floorplan depth and construct image-side ray observations. From the gravity-aligned image and mask, the system predicts floorplan depth and converts it into equiangular image-side rays. Unlike F3Loc, FreeLoc queries floorplan-side rays directly from the geometric floorplan rather than retrieving them from a precomputed pose-indexed database.

It uniformly samples a compact set of pose hypotheses and evaluates their ray descriptors with a rotation-aware ray kernel. For each candidate position and yaw, the kernel casts equiangular rays and computes ray distances by analytic boundary intersection. The query uses the native line-segment structure of floorplans and computes analytic ray-boundary intersections directly, rather than dense hit-testing over rasterized occupancy grids.

This reduces query cost while preserving accurate geometric ray measurements. A vectorized batched GPU operation evaluates candidate poses and boundary segments in parallel. Together with compact pose sampling, this enables efficient online floorplan ray extraction for real-time coarse retrieval.

After coarse retrieval, the top-K anchors are plausible but still discretized pose hypotheses. The diffusion-aided pose refinement module formulates refinement as residual pose optimization and produces accurate continuous pose estimates beyond the coarse sampling resolution.

For each coarse candidate, the system estimates a relative correction conditioned on image-side and corresponding floorplan-side ray observations. Residual pose estimation is a conditional diffusion process rather than one-shot regression, with iterative denoising that progressively refines the correction under paired ray guidance.

The image-side and floorplan-side rays are first encoded with a shared PointNet-style ray encoder. A Ray Discrepancy Encoder computes soft correspondences between paired ray features and aggregates them into a global conditioning feature. The conditional denoiser takes the noisy residual, diffusion timestep, and conditioning feature as input, and is trained to predict the injected noise.

For sequential RGB observations, the goal is temporally consistent pose estimates. FreeLoc constructs observation likelihoods online from sparse coarse-to-fine outputs: coarse candidates provide broad pose-space support, while refined hypotheses sharpen high-confidence regions.

The resulting likelihood is fused over time with a histogram filter. Evaluation uses Gibson and Structured3D, two widely used datasets with different distributions. Gibson contains 118 indoor scenes with floorplans and monocular images captured with a one-hundred-eight-degree field of view.

Its single-frame split has 24,779 four-view sequences, while its sequential split has 118 trajectories ranging from 280 to 5,152 steps. Structured3D contains 3,296 furnished indoor environments and 78,453 perspective images captured with an eighty-degree horizontal field of view.

The comparison includes F3Loc as the primary depth-based baseline, along with PF-Net and LASER as representative non-depth-based methods. For single-frame localization, the evaluation reports recall at translation thresholds of zero point one, zero point five, and one point zero meters, plus a joint threshold of one meter and thirty degrees.

For sequential localization, it reports success rate and tracking accuracy; a trajectory succeeds when predictions stay within the threshold over the last ten frames. On both Gibson(f) and Structured3D, FreeLoc consistently outperforms all baselines.

Compared with F3Loc, it improves all recall metrics by 4.7 to 8.7 percentage points on Gibson(f), and by 1.4 to 8.9 percentage points on Structured3D. The reported improvements are attributed mainly to the pose refinement stage.

Because refinement optimizes coarse anchors in continuous pose space, predictions can move beyond sampled locations and achieve finer pose accuracy. Looking across the Gibson(f) columns, the “Ours” row reaches 10.0 at zero point one meter and 44.6 at one meter, while F3Loc shows 5.3 and 37.9.

The Structured3D block shows the same separation: 3.1 versus 1.7 at zero point one meter, and 30.5 versus 23.0 at one meter. On Gibson(t), with sequence length set to 100 frames, FreeLoc achieves the best success rate and overall tracking accuracy. Compared with F3Loc, it improves success rate by 13.5 percentage points and reduces RMSE overall by 82.8 percent.

For successfully localized sequences, RMSE remains comparable: 0.15 meters for FreeLoc versus 0.14 meters for F3Loc. The slight gap is mainly attributed to reconstructing dense observation likelihoods from sparse online candidates, which may reduce local precision compared with dense likelihood evaluation over a precomputed ray database.

When coarse hypotheses lie within 0.5, 1.0, and 1.5 meters of ground truth, diffusion refinement consistently reduces average translation error. When all hypotheses are included, improvement is marginal, from 3.14 meters to 3.09 meters.

The limited gain occurs because the all set includes poorly initialized and cross-room hypotheses that lie outside the effective range of local residual correction. The intended division of labor is clear: online coarse retrieval narrows the global search, while diffusion refinement performs precise local residual correction.

Reading across each row, the “After” column is lower than the corresponding “Before” entry: the gaps are seven hundredths of a meter for the first row, one tenth of a meter for the second, and one tenth of a meter for the third. The “All” row, by contrast, shows only a five-hundredths-of-a-meter gap between its two columns.

FreeLoc treats the floorplan as a directly queryable geometric map rather than a dense scene-specific offline database. Its online coarse-to-fine pipeline combines on-the-fly floorplan ray querying with diffusion-aided pose refinement to retrieve plausible anchors and refine them into accurate continuous pose estimates.

The same anchors and refined hypotheses construct online observation likelihoods for sequential localization, bridging single-frame localization and probabilistic temporal fusion without offline databases. Experiments demonstrate accurate and efficient localization, along with immediate deployability in new environments.

Figure 3 illustrates FreeLoc operating in the HGE building using a narrow-field-of-view handheld iPhone sequence. The center floorplan traces predicted locations in blue alongside ground truth in green, while the surrounding panels pair time-ordered RGB observations with posterior probability maps.

The initially diffuse posterior becomes spatially concentrated as observations accumulate, showing how the system tracks a trajectory through rooms and corridors without scene-specific preprocessing. FreeLoc replaces dense offline lookup with an online coarse-to-fine pipeline: floorplan rays find plausible anchors, diffusion refines them continuously, and the same evidence supports sequential tracking.

The result is stronger localization without a scene-specific database.