From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization without Ray Matching
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "From Uncertainty to Determinism".
Dev: Here's the gist of the paper. Visual Floorplan Localization (FLoc) has emerged as a promising solution for indoor localization by matching egocentric images against minimalist structural maps.
Rosa: This paper proposes a coarse-to-fine visual FLoc framework that progresses from uncertainty to determinism, bypassing intermediate ray-matching paradigms.
Dev: Problem Formulation and Motivation Visual FLoc is formulated as estimating the continuous conditional pose distribution p(p I, M) rather than deterministic regression.
Rosa: The optimal pose is theoretically defined by the maximum a posteriori (MAP) estimation of Eq.
Title and authors: Rosa: So we're looking at this paper called "From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization without Ray Matching". It tackles how you can figure out where a camera is in a building by matching what it sees to the floorplan map.
Dev: It seems they are focusing on moving away from those old ray matching methods that generate sparse lines, because those methods create this information bottleneck before you even get to the actual localization part.
Taro: So instead of predicting these geometric rays, this paper proposes a different approach where the system directly estimates the entire continuous conditional pose distribution over the map, which is written as p of p given I and M.
Rosa: That's right, they are moving from trying to find one single best answer to modeling all possible answers simultaneously before refining them later.
Dev: They suggest a coarse-to-fine framework where the first stage uses a generative model over the entire floorplan to get a set of plausible candidate poses, which they call p k c.
Taro: And then in the second stage, they take those candidates and use local information to refine them into more precise sub-meter pose residuals.
Rosa: That sounds like they are tackling the problem by first getting a broad idea of where you might be before zooming in on the exact spot using local details.
Dev: The authors say this avoids the issue where visual observations that look identical might map to different locations because it doesn't force a single, deterministic ray prediction.
Taro: They mention that they use a conditional denoising network during the first stage to model this multimodal distribution, injecting Gaussian noise into the pose representation over time.
Rosa: It sounds like they are using diffusion models to explore that entire range of possibilities across the whole map at once.
Dev: The core idea is to bypass those heavy offline preprocessing steps that ray matching needs for its databases, because they propose a rendering-free generative localization framework called CF2Loc.
Taro: And the results on their benchmarks, like Sthree dee and ZInD, show they achieve better recall figures compared to older methods, specifically improving the zero point five meter recall from eleven point one percent down to forty-five point five percent on ZInD.
Rosa: That jump is significant because it shows that this method is more robust when dealing with those tricky cross-modal asymmetries in indoor scenes.
Dev: They also found that applying joint map-pose rotation augmentation really helped boost the recall numbers, pushing the zero point five meter recall up to fifty-eight point five percent and the one point zero meter recall to seventy three point two percent on those benchmarks.
Taro: The ablation studies confirm they aren't just averaging all those diffusion samples randomly; they have to preserve those dominant mode hypotheses in the distribution, otherwise it won't work well.
Rosa: And integrating that local refiner was crucial for getting that fine-grained accuracy, especially once they enabled local patch canonicalization for the crops.
Dev: The final result is that this approach provides state-of-the-art accuracy while also supporting real-time inference without needing any offline map preprocessing or test time look up tables.
Taro: It’s a scalable alternative to those traditional ray matching paradigms, which rely on dense precomputed databases for every map.
Rosa: So, to wrap up the paper "From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization without Ray Matching", we see they've successfully moved from uncertainty to determinism by using a generative model first and then local refinement.
Dev: This framework parameterizes the entire multimodal pose posterior via an image conditioned pose diffusion model, which is then followed by a lightweight local refiner for precise residual regression.
Taro: The implication for autonomy is that you get a system that can handle the ambiguity of indoor environments by exploring multiple plausible global hypotheses and then snapping into place locally.
Rosa: This work offers a way to localize indoors without needing those resource-intensive ray matching databases, which is pretty important when you don't have time to prepare huge sets of data for every single map.
Dev: It really shifts the focus from deterministic regression to estimating the continuous conditional pose distribution, making the localization problem itself probabilistic and more realistic.
Taro: I think the big win here is that it shows how you can build a system that progresses from a very fuzzy guess across the whole room to a very precise measurement using localized floorplan crops.
Rosa: It sounds like they've given us a way to do this in real-time, which is what we need for actual robotics outside of controlled lab settings.
Dev: And while they claim it works without offline preprocessing, you have to remember that the performance still depends on how well that initial global proxy distribution captures the true multimodal nature of the scene.
Taro: That's a fair point; if the first stage misses a dominant mode, the refinement stage won't be able to fix it correctly.
Rosa: So, looking ahead at this paper, I think we should keep watching how they handle those real-world misbehaves when they move this out of the controlled environment.
Dev: And for now, this CF2Loc framework gives us a solid foundation that doesn't rely on precomputed structures for every floorplan.
Taro: We'll see if the system can gracefully recover when things get truly messy outside of the perfect test conditions they used in their evaluations.
The paper's summary: Rosa: So, we're talking about how this new visual floorplan localization system works by moving away from those old ray matching methods that create sparse lines in the first place.
Dev: It’s about modeling the entire continuous conditional pose distribution across a whole map directly, instead of just picking one single best answer.
Rosa: Right, so they're using this coarse-to-fine structure where you get a broad idea of where you might be across the whole floorplan first.
Dev: And then you use local information to refine those broad guesses into very precise sub-meter corrections by looking at small patches of the map.
Rosa: It means they bypass that information bottleneck before localization and they don't need those huge offline databases that ray matching requires for every single map.
Dev: The results on benchmarks like Sthree dee and ZInD are pretty strong, showing improvements in recall figures, like taking the five point one percent recall up to forty-five point five percent on ZInD.
Rosa: And they found that adding joint map and pose rotation augmentation helped push that recall even higher, up to fifty-eight point five percent for the small distance guesses.
Dev: That local refiner is key because it eliminates those structural ambiguities locally, which gives them better fine-grained accuracy without needing to do any test time lookups.
Rosa: It really shows how you can build a system that handles the ambiguity of indoor scenes by exploring multiple plausible global hypotheses and then snapping into place locally.
Dev: This approach shifts the problem from just predicting a single pose to estimating the entire multimodal posterior, which is much more realistic for real-world scenarios where things aren't perfectly clear.
Rosa: It’s interesting because it doesn't rely on those heavy preprocessing steps or those test time lookup tables that ray matching needs.
Dev: The challenge then becomes making sure that initial global proxy distribution actually captures the true multimodal nature of the scene, otherwise the refinement stage won't be able to fix it correctly.
Rosa: So, what this means for us is a scalable way to localize indoors without needing those resource-intensive ray matching databases for every single map.
Dev: And this moves the focus from deterministic regression to estimating that continuous conditional pose distribution, making the localization problem itself probabilistic and more realistic.
Rosa: We'll see how this performs when it gets out of the controlled lab setting, especially with those real-world misbehaves Taro is testing for.
The paper's improvements: Tom: So, we're looking at how they're changing the approach by focusing on direct conditional pose distribution estimation instead of those old ray representations.
Rosa: That’s right, they're bypassing that information bottleneck before localization and they aren't using those huge offline databases that ray matching requires for every single map.
Dev: They introduce this coarse-to-fine structure where you get a broad idea of where you might be across the whole floorplan first.
Rosa: And then you use local information to refine those broad guesses into very precise sub-meter corrections by looking at small patches of the map.
Dev: The local refiner is crucial because it eliminates those structural ambiguities locally, which gives them better fine-grained accuracy without needing to do any test time lookups.
Rosa: It means they’re not just guessing a single point; they're modeling all the possibilities and then zooming in on the right one using local context.
Dev: This method parameterizes that entire multimodal posterior via an image conditioned pose diffusion model, which is then followed by that lightweight local refiner for precise residual regression.
Rosa: It really shows how you can build a system that handles the ambiguity of indoor scenes by exploring multiple plausible global hypotheses and then snapping into place locally.
Dev: This approach shifts the problem from just predicting a single pose to estimating the entire multimodal posterior, which is much more realistic for real-world scenarios where things aren't perfectly clear.
Rosa: It’s interesting because it doesn't rely on those heavy preprocessing steps or those test time lookup tables that ray matching needs.
Dev: The challenge then becomes making sure that initial global proxy distribution actually captures the true multimodal nature of the scene, otherwise the refinement stage won't be able to fix it correctly.
Rosa: So, what this means for us is a scalable way to localize indoors without needing those resource-intensive ray matching databases for every single map.
Dev: This moves the focus from deterministic regression to estimating that continuous conditional pose distribution, making the localization problem itself probabilistic and more realistic.
Rosa: We'll see how this performs when it gets out of the controlled lab setting, especially with those real-world misbehaves Taro is testing for.
Conclusion: Rosa: So, to wrap up "From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization without Ray Matching," they've shown how you can go from just guessing a general location to a very precise measurement by using this generative framework first and then local refinement.
Dev: It’s interesting because it completely bypasses the need for those heavy offline preprocessing steps or test time lookups that ray matching requires for every single map.
Rosa: Exactly, so you get this system that can handle the ambiguity of indoor scenes by exploring multiple plausible global hypotheses and then snapping into place locally.
Dev: This moves the focus from just predicting a single pose to estimating that continuous conditional pose distribution, which is much more realistic for real-world scenarios where things aren't perfectly clear.
Rosa: We'll see how this performs when it gets out of the controlled lab setting, especially with those real-world misbehaves Taro is testing for.
Dev: The challenge then becomes making sure that initial global proxy distribution actually captures the true multimodal nature of the scene, otherwise the refinement stage won't be able to fix it correctly.
Rosa: It’s a scalable way to localize indoors without needing those resource-intensive ray matching databases for every single map.
Dev: This moves the focus from deterministic regression to estimating that continuous conditional pose distribution, making the localization problem itself probabilistic and more realistic.
Rosa: We'll see how this performs when it gets out of the controlled lab setting, especially with those real-world misbehaves Taro is testing for.
Dev: The safety case here is pretty strong because it handles uncertainty directly in the modeling stage instead of relying on a single deterministic guess.
Rosa: It sounds like a solid foundation that doesn't rely on precomputed structures for every floorplan, which is pretty important when you don't have time to prepare huge sets of data for every single map.
Dev: And if we can get this kind of direct modeling working reliably in real-time, it opens the door for more robust systems in mobile robotics where the environment is constantly changing.
Rosa: Yeah, I think this paper provides a really tangible method for how we can tackle that continuous pose distribution problem without being stuck with the limitations of older ray matching techniques.
Dev: Exactly, and we'll see if this system can gracefully recover when things get truly messy outside of the perfect test conditions they used in their evaluations.
Taro: I think what really stands out is how they address those real-world misbehaves during testing; it shows the system can handle uncertainty more gracefully than just averaging samples randomly.
Rosa: That’s true, and seeing how it performs when the world misbehaves outside of a lab setting is what we need to see next.
Dev: For now, this CF2Loc framework gives us a solid foundation that doesn't rely on precomputed structures for every floorplan.
Shiyong Meng, *Bolei Chen*, *Ping Zhong*, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
cs.RO, cs.CV
Submitted: 2026-07-29
Updated: 2026-11-16
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Visual Floorplan Localization (FLoc) has emerged as a promising solution for indoor localization by matching egocentric images against minimalist structural maps.
Key concepts
- Continuous Conditional Pose Distribution
- Instead of guessing a single location, this concept models the entire range of possible poses as a continuous distribution. It represents all plausible locations for the camera based on both the visual input and the floorplan map, allowing for more robust uncertainty modeling than traditional methods.
- Global Pose Diffusion Model
- This is a generative model that uses progressive noise injection to estimate how a pose should look across the whole map. It learns to generate a diverse set of possible poses simultaneously, capturing the multimodal nature of localization without relying on intermediate ray calculations.
- Local Pose Refiner
- This component takes coarse guesses from the global model and then focuses only on small sections (crops) of the floorplan. It predicts precise, bounded errors (residuals) for translation and orientation within that local area, significantly improving accuracy by handling fine details.
- Coarse-to-Fine Inference
- This is a two-stage process where the system first generates broad hypotheses across the whole map (coarse stage) and then uses those hypotheses to refine them locally (fine stage). This hierarchical approach ensures both global context and precise local accuracy are captured effectively.
Terminology
Summary
Visual Floorplan Localization (FLoc) has emerged as a promising solution for indoor localization by matching egocentric images against minimalist structural maps. p. 0 cn, jxwang@mail. cn
Abstract
Visual Floorplan Localization (FLoc) has emerged as a
promising solution for indoor localization by matching egocentric images against minimalist structural maps. Existing
ray-matching-based methods tackle this by explicitly predicting sparse geometric or semantic rays, which inherently incurs
information loss and demands resource-intensive preprocessing alongside exhaustive matching during inference. In this
paper, we bypass the intermediate ray-matching paradigm
and propose a coarse-to-fine visual FLoc framework that progresses from uncertainty to determinism.Where the paper says it
Where the paper says it
Problem Formulation and Motivation
Visual FLoc is formulated as estimating the continuous conditional pose distribution p(p I, M) rather than deterministic regression. The optimal pose is theoretically defined by the maximum a posteriori (MAP) estimation of Eq. p. 2 To accommodate the intrinsic ambiguities of
indoor structures, we move beyond deterministic regression
and instead formulate localization as estimating the continuous conditional pose distribution p(p | I, M). The optimal
pose is theoretically defined by the maximum a posteriori
(MAP) estimation:
p
∗ = arg max
p
p(p | I, M). Motivation. The pronounced cross-modal asymmetry between feature-rich visual observations and minimalist 2D
floorplans renders visual localization a highly ill-posed problem. Existing paradigms struggle with this multimodal nature. Deterministic regressors
enforce a single-peak constraint on the output, leading to
mode collapse and significant averaging errors in ambiguous
regions.Where the paper says it
Where the paper says it
Where the paper says it
Global Pose Diffusion
The global pose diffusion model directly estimates the continuous conditional pose distribution across the entire floorplan. This involves modeling the multimodal pose distribution through conditional diffusion, where Gaussian noise is progressively injected into the pose representation zt. p. 2 Following the standard DDPM (Ho, Jain, and Abbeel
2020) formulation, Gaussian noise is progressively injected
into the pose representation:
zt =
√
α¯tz0 +
√
1 − α¯tϵ, ϵ ∼ N (0, I), (5)
where z0 = p˜. A conditional denoising network predicts the injected
noise from the noisy pose, diffusion timestep, global image
descriptor, and observation-conditioned map tokens:
ˆϵ = ϵθ(zt, t, FˆM, gI ). Following the standard DDPM (Ho, Jain, and Abbeel
2020) formulation, Gaussian noise is progressively injected
into the pose representation:
zt =
√
α¯tz0 +
√
1 − α¯tϵ, ϵ ∼ N (0, I), (5)
where z0 = p˜. A conditional denoising network predicts the injected
noise from the noisy pose, diffusion timestep, global image
descriptor, and observation-conditioned map tokens:
ˆϵ = ϵθ(zt, t, FˆM, gI ).Where the paper says it
Where the paper says it
Local Pose Refinement
The local refiner is introduced to predict bounded sub-meter pose residuals from candidate-centered floorplan crops. This refinement operates on localized floorplan crops while reusing the visual tokens extracted during the coarse stage. p. 3 2(a), the refiner operates on localized floorplan crops while reusing the visual tokens extracted during
the coarse stage. Candidate-Centered Local Refinement. (9)
Using the fused local representation, three lightweight prediction heads (unified as Rϕ) respectively estimate position
logits l
k = {l
k
j
}
Nc
j=1, local heading vectors Vk = {v
k
j
}
Nc
j=1,
and a candidate confidence score sˆ
k
:
(l
k
, V
k
, sˆ
k
) = Rϕ(I, Mk
c , p
k
c ). (10)
Rather than directly regressing coordinates via global feature pooling, we compute translation residuals by constructing a spatial localization distribution over local map tokens:
π
k
j =
exp(l
k
j /τ )
P
n
exp(l
k
n/τ )
, ∆xˆ
k =
X
j
π
k
j uj , (11)
where uj denotes the metric coordinate of token j in the
candidate-local frame. Similarly, spatial heading vectors are
aggregated via probability-weighted summation to determine
the rotation residual:
∆ˆθ
k = atan2
"X
j
π
k
j v
k
j
#
1
,
"X
j
π
k
j v
k
j
#
2
. (12)
The refined pose p
k
r
is obtained by applying the predicted
residual to the coarse candidate pose p
k
c ⊕ (∆xˆ
k
, ∆ˆθ
k
).Where the paper says it
Where the paper says it
Where the paper says it
Coarse-to-Fine Inference and Results
During inference, N pose particles are initialized from Gaussian noise and refined through 10 reverse diffusion steps to approximate the continuous multimodal pose posterior. Candidate modes are extracted via Kernel Density Estimation (KDE) to form coarse candidate modes. p. 4 To extract representative pose modes from the sample
set, we estimate particle density over the pose manifold via
Kernel Density Estimation (KDE):
D(p) = 1
N
XN
i=1
KH(p − p
i
), (17)
where KH(·) denotes a product kernel in SE(2) space with
bandwidth matrix H. The K density local maxima are extracted as coarse candidate modes {p
k
c }
K
k=1. Finally, candidate hypotheses are
re-ranked based on their predicted confidence scores:
k
∗ = arg max
k∈{1,. ,K}
sˆ
k
, (18)
and the top-ranked candidate p
k
∗
r
is output as the final localization pose. Our method effectively balances global multi-hypothesis tracking and local
sub-meter refinement without requiring any offline map preprocessing or test-time lookup tables. Comprehensive results
on the S3D (full) and ZInD benchmarks demonstrate that our
approach achieves state-of-the-art accuracy and robustness. When training with semantic floorplans, our method continues to establish new SOTA performance. Compared with
the strongest semantic baseline, FoD, our framework improves recall on S3D (full) from 12. 2024; Grader and Averbuch-Elor 2025;
Ye et al. 2026), we report recall metrics at spatial accuracy
thresholds of 0. 1999) is the most popular framework for 2D localization on pure geometry maps. (3) LASER (Min et al. Our method effectively balances global multi-hypothesis tracking and local
sub-meter refinement without requiring any offline map preprocessing or test-time lookup tables. Comprehensive results
on the S3D (full) and ZInD benchmarks demonstrate that our
approach achieves state-of-the-art accuracy and robustness.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Ablation Studies
Ablation studies confirm the necessity of preserving dominant mode hypotheses in a multimodal pose distribution rather than naively averaging diffusion samples. Applying joint map-pose rotation augmentation further boosts the 0. p. 6 This confirms the necessity of preserving dominant mode hypotheses in a multimodal pose distribution rather than naively averaging diffusion samples. Applying joint map-pose rotation augmentation further boosts
the 0. 2024; Grader and Averbuch-Elor 2025;
Ye et al. 2026), we report recall metrics at spatial accuracy
thresholds of 0. 2024; Grader and Averbuch-Elor 2025;
Ye et al. 2026), we report recall metrics at spatial accuracy
thresholds of 0. 2%,
indicating enhanced robustness to arbitrary global floorplan
orientations. Integrating the local refiner consistently improves fine-grained localization accuracy, with gains becoming significantly more pronounced after enabling local patch
canonicalization (rotation alignment). Parametric Studies
We conduct sensitivity analyses on the local map size. As
shown in Table 4, a 5 m × 5 m crop yields the best accuracy by balancing spatial layout context with local unimodality—smaller patches lack context, whereas larger ones
reintroduce structural ambiguities.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
In this paper, we revisit visual floorplan localization (FLoc) from the perspective of direct conditional pose distribution estimation and propose a rendering-free, coarse-to-fine generative framework. Our method parameterizes the full multimodal pose posterior via an image-conditioned pose diffusion model, followed by a lightweight local refiner for precise residual regression. p. 7 Conclusion
In this paper, we revisit visual floorplan localization (FLoc)
from the perspective of direct conditional pose distribution estimation and propose a rendering-free, coarse-to-fine
generative framework. Rather than relying on intermediate
1D ray prediction or exhaustive observation matching, our
method parameterizes the full multimodal pose posterior via
an image-conditioned pose diffusion model, followed by a
lightweight local refiner for precise residual regression. This
coarse-to-fine formulation seamlessly bridges global uncertainty modeling with local unimodal determinism, enabling
robust multi-hypothesis tracking and fine-grained pose estimation within a unified architecture applicable to both
geometric and semantic floorplans. Extensive evaluations
on the S3D (full) and ZInD benchmarks demonstrate that
our approach consistently establishes new SOTA accuracy
while supporting real-time inference without requiring offline map preprocessing. Extensive evaluations
on the S3D (full) and ZInD benchmarks demonstrate that
our approach consistently establishes new SOTA accuracy
while supporting real-time inference without requiring offline map preprocessing. We hope this work provides a scalable, rendering-free alternative to traditional ray-matching
paradigms and inspires further research in generative crossmodal spatial reasoning.Where the paper says it
Where the paper says it
Where the paper says it
References
Arandjelović, R.; Gronat, P. p. 7 References
Arandjelović, R. ; Gronat, P. ; Gronat, P. ; Torii, A. ; Torii, A. ; Pajdla, T. ; Pajdla, T. ; and Sivic, J. ; and Sivic, J. NetVLAD: CNN Architecture for Weakly Supervised
Place Recognition. NetVLAD: CNN Architecture for Weakly Supervised
Place Recognition. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 5297–
5307.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Balntas, V. p. 0 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. ; and Prisacariu, V. RelocNet:
Continuous Metric Learning Relocalisation Using Neural
Nets. RelocNet:
Continuous Metric Learning Relocalisation Using Neural
Nets. In Proceedings of the European Conference on Computer Vision (ECCV), 751–767.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Brachmann, E. p. 1 2019; Sattler, Leibe, and Kobbelt 2017), or direct pose regression using deep networks (Kendall, Grimes,
and Cipolla 2015; Walch et al. 2017; Brachmann et al. Brachmann, E. ; Krull, A. ; Krull, A. ; Nowozin, S. ; Nowozin, S. ; Shotton, J. ; Shotton, J. ; Michel,
F. ; Michel,
F. ; Gumhold, S. ; Gumhold, S. ; and Rother, C. ; and Rother, C. Dsac-differentiable
ransac for camera localization. Dsac-differentiable
ransac for camera localization. In Proceedings of the IEEE
conference on computer vision and pattern recognition,
6684–6692.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Chen, B.; Kang, J.; Yang, H.; Zhong, P. p. 0 From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. ; and Wang, J. Perspective from a Higher Dimension: Can 3D Geometric
Priors Help Visual Floorplan Localization? Perspective from a Higher Dimension: Can 3D Geometric
Priors Help Visual Floorplan Localization? In Proceedings
of the 33rd ACM International Conference on Multimedia
(ACM MM), 7181–7190.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Chen, B.; Yan, S.; Cui, Y.; Kang, J.; Zhong, P.; and Wang, J. 2026a. Perspective from a Broader Context: Can Room Style Knowledge Help Visual Floorplan Localization? p. 7 ; and Wang,
J. Perspective from a Broader Context: Can Room
Style Knowledge Help Visual Floorplan Localization? Perspective from a Broader Context: Can Room
Style Knowledge Help Visual Floorplan Localization? In
Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 2787–2795.Where the paper says it
Where the paper says it
Chen, C.; Wang, R.; Vogel, C. p. 7 In
Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 2787–2795. ; Vogel, C. ; Vogel, C. ; and Pollefeys, M. ; and Pollefeys, M. F
3Loc: Fusion and Filtering for Floorplan Localization. F
3Loc: Fusion and Filtering for Floorplan Localization. In
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 18029–18038.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Chen, K.; Huang, Y. p. 0 Compared with dense 3D models, floorplans are lightweight, easy to maintain, and readily available
in built environments, making visual FLoc particularly appealing for autonomous robot navigation (Li et al. 2025;
Huang et al. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. ;
and Wang, L. FloorPlan-VLN: A New Paradigm for
Floor Plan Guided Vision-Language Navigation. FloNa: Floor Plan Guided Embodied Visual Navigation. p. 7 ; and Liu, F. p. 0 Introduction
Visual Floorplan Localization (FLoc) (Chen et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. FloNa: Floor Plan Guided Embodied Visual Navigation. In Proceedings of the AAAI Conference on Artificial
Intelligence, volume 39, 14610–14618.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Cho, J.; Cai, R.; and Averbuch-Elor, H. p. 1 2026; Chen et al. 2025, 2026a; Grader and Averbuch-Elor
2025; Ye et al. 2026; Chen et al. 2025, 2026a; Grader and Averbuch-Elor
2025; Ye et al. ; and Averbuch-Elor, H. SceneAligner:
3D-Grounded Floorplan Localization in the Wild. SceneAligner:
3D-Grounded Floorplan Localization in the Wild. arXiv
preprint arXiv:2605.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Cruz, S.; Hutchcroft, W. p. 7 arXiv
preprint arXiv:2605. ; Hutchcroft, W. ; Hutchcroft, W. ; Khosravan, N. ; Hutchcroft, W. ; Khosravan, N. ; Khosravan, N. ; Boyadzhiev,
I. ; Boyadzhiev,
I. ; and Kang, S. ; and Kang, S. Zillow Indoor Dataset: Annotated
Floor Plans With 360° Panoramas and 3D Room Layouts.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Dellaert, F. p. 4 Its observation model
aims to learn the similarity between an image and the corresponding map patch. (2) MCL (Dellaert et al. Dellaert, F. ; Burgard, W. ; Burgard, W. ; and Thrun, S. ; and Thrun, S. Monte
Carlo Localization for Mobile Robots. Monte
Carlo Localization for Mobile Robots. In Proceedings of the
IEEE International Conference on Robotics and Automation
(ICRA), volume 2, 1322–1328.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Grader, Y.; and Averbuch-Elor, H. 2025. Supercharging Floorplan Localization with Semantic Rays. p. 7 ; and Averbuch-Elor, H. Supercharging
Floorplan Localization with Semantic Rays. Supercharging
Floorplan Localization with Semantic Rays. In Proceedings
of the IEEE/CVF International Conference on Computer Vision (ICCV), 27116–27125.Where the paper says it
Where the paper says it
Ho, J.; Jain, A.; and Abbeel, P. p. 2 To circumvent boundary discontinuities associated with
angular regression, the camera pose is parameterized as
p˜ = (x, y, cos θ,sin θ), where orientation is recovered using atan2(·) during inference. Following the standard DDPM (Ho, Jain, and Abbeel
2020) formulation, Gaussian noise is progressively injected
into the pose representation:
zt =
√
α¯tz0 +
√
1 − α¯tϵ, ϵ ∼ N (0, I), (5)
where z0 = p˜. ; and Abbeel, P. Denoising Diffusion Probabilistic Models. Denoising Diffusion Probabilistic Models. In Larochelle, H. In Larochelle, H. ; Ranzato, M. ; Ranzato, M. ;
Hadsell, R. ;
Hadsell, R. ; Balcan, M. ; Balcan, M. ; and Lin, H. ; and Lin, H. , Advances in Neural Information Processing Systems, volume 33, 6840–6851.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Howard-Jenkins, H. p. 1 End-to-End Visual FLoc
End-to-end visual FLoc directly learns a mapping from
egocentric observations and floorplans to camera poses
without explicit geometric matching. Representative methods, such as LaLaLoc (Howard-Jenkins, Ruiz-Sarmiento,
and Prisacariu 2021), LaLaLoc++ (Howard-Jenkins and
Prisacariu 2022), and PF-Net (Karkus, Hsu, and Lee 2018),
learn joint image–floorplan representations for direct pose
regression or particle-filter-based localization. End-to-End Visual FLoc
End-to-end visual FLoc directly learns a mapping from
egocentric observations and floorplans to camera poses
without explicit geometric matching. Representative methods, such as LaLaLoc (Howard-Jenkins, Ruiz-Sarmiento,
and Prisacariu 2021), LaLaLoc++ (Howard-Jenkins and
Prisacariu 2022), and PF-Net (Karkus, Hsu, and Lee 2018),
learn joint image–floorplan representations for direct pose
regression or particle-filter-based localization. ; and Prisacariu, V. LaLaLoc++: Global Floor Plan Comprehension for Layout
Localisation in Unvisited Environments. LaLaLoc++: Global Floor Plan Comprehension for Layout
Localisation in Unvisited Environments. In Proceedings of
the European Conference on Computer Vision (ECCV), 693–
709.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Howard-Jenkins, H.; Ruiz-Sarmiento, J. p. 1 End-to-End Visual FLoc
End-to-end visual FLoc directly learns a mapping from
egocentric observations and floorplans to camera poses
without explicit geometric matching. Representative methods, such as LaLaLoc (Howard-Jenkins, Ruiz-Sarmiento,
and Prisacariu 2021), LaLaLoc++ (Howard-Jenkins and
Prisacariu 2022), and PF-Net (Karkus, Hsu, and Lee 2018),
learn joint image–floorplan representations for direct pose
regression or particle-filter-based localization. ; and Prisacariu,
V. LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments. LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments. In Proceedings of the
IEEE/CVF International Conference on Computer Vision
(ICCV), 10107–10116.Where the paper says it
Where the paper says it
Where the paper says it
Huang, W.; Li, J.; Wan, Z.; Di, H.; Liang, W. p. 7 In Proceedings of the
IEEE/CVF International Conference on Computer Vision
(ICCV), 10107–10116. ; Liang, W. In Proceedings of the
IEEE/CVF International Conference on Computer Vision
(ICCV), 10107–10116. ; Liang, W. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. ; and
Yang, Z. Floor Plan-Guided Visual Navigation Incorporating Depth and Directional Cues. Floor Plan-Guided Visual Navigation Incorporating Depth and Directional Cues. arXiv preprint
arXiv:2511.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Karkus, P.; Hsu, D.; and Lee, W. p. 1 End-to-End Visual FLoc
End-to-end visual FLoc directly learns a mapping from
egocentric observations and floorplans to camera poses
without explicit geometric matching. Representative methods, such as LaLaLoc (Howard-Jenkins, Ruiz-Sarmiento,
and Prisacariu 2021), LaLaLoc++ (Howard-Jenkins and
Prisacariu 2022), and PF-Net (Karkus, Hsu, and Lee 2018),
learn joint image–floorplan representations for direct pose
regression or particle-filter-based localization. ; and Lee, W. Particle Filter
Networks with Application to Visual Localization. Particle Filter
Networks with Application to Visual Localization. In Proceedings of the 2nd Conference on Robot Learning (CoRL),
volume 87 of Proceedings of Machine Learning Research,
169–178.Where the paper says it
Where the paper says it
Where the paper says it
Kendall, A. p. 1 2016; Balntas, Li, and Prisacariu 2018), 2D–3D
feature matching with Structure-from-Motion (SfM) models
(Sarlin et al. 2019; Sattler, Leibe, and Kobbelt 2017), or direct pose regression using deep networks (Kendall, Grimes,
and Cipolla 2015; Walch et al. 2016; Balntas, Li, and Prisacariu 2018), 2D–3D
feature matching with Structure-from-Motion (SfM) models
(Sarlin et al. 2019; Sattler, Leibe, and Kobbelt 2017), or direct pose regression using deep networks (Kendall, Grimes,
and Cipolla 2015; Walch et al. 2016; Balntas, Li, and Prisacariu 2018), 2D–3D
feature matching with Structure-from-Motion (SfM) models
(Sarlin et al. 2019; Sattler, Leibe, and Kobbelt 2017), or direct pose regression using deep networks (Kendall, Grimes,
and Cipolla 2015; Walch et al. ; and Cipolla, R. Posenet: A
convolutional network for real-time 6-dof camera relocalization. Posenet: A
convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference
on computer vision, 2938–2946.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Li, J.; Huang, W. p. 0 Compared with dense 3D models, floorplans are lightweight, easy to maintain, and readily available
in built environments, making visual FLoc particularly appealing for autonomous robot navigation (Li et al. 2025;
Huang et al.Where the paper says it
Liu, L.; Li, H.; and Dai, Y. p. 0 Introduction
Visual Floorplan Localization (FLoc) (Chen et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. ; and Dai, Y. Efficient Global 2D–
3D Matching for Camera Localization in a Large-Scale 3D In Proceedings of the IEEE International Conference
on Computer Vision (ICCV), 2372–2381. Loshchilov, I. p. 5 All attention mechanisms employ four heads with a dropout rate of 0. The diffusion model is optimized via AdamW (Loshchilov
and Hutter 2017) using a cosine noise schedule (T = 1000),
a batch size of 8, an initial learning rate of 10−4
, a minimum
learning rate of 10−5
, and a weight decay of 10−4
.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Loshchilov, I.; and Hutter, F. p. 5 All attention mechanisms employ four heads with a dropout rate of 0. The diffusion model is optimized via AdamW (Loshchilov
and Hutter 2017) using a cosine noise schedule (T = 1000),
a batch size of 8, an initial learning rate of 10−4
, a minimum
learning rate of 10−5
, and a weight decay of 10−4
. ; and Hutter, F. Decoupled Weight Decay
Regularization. Decoupled Weight Decay
Regularization. In International Conference on Learning
Representations.Where the paper says it
Where the paper says it
Where the paper says it
Min, Z.; Khosravan, N.; Bessinger, Z. p. 8 ; Khosravan, N. ; Bessinger, Z. ; Bessinger, Z. ; Narayana, M. ; Khosravan, N. ; Boyadzhiev,
I. ; and Boyadzhiev, I. LASER: LAtent
SpacE Rendering for 2D Visual Localization. LASER: LAtent
SpacE Rendering for 2D Visual Localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 11122–11131.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Sarlin, P.-E.; Cadena, C. p. 8 In Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), 11122–11131. ; Cadena, C. ; Cadena, C. ; Siegwart, R. ; Siegwart, R. ; and Dymczyk, M. ; and Dymczyk, M. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),
12716–12725.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Sattler, T. p. 0 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. ; and Kobbelt, L. Efficient and
Effective Prioritized Matching for Large-Scale Image-Based
Localization. Efficient and
Effective Prioritized Matching for Large-Scale Image-Based
Localization. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 39(9): 1744–1756.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Vaswani, A. p. 5 The floorplan encoder comprises a
convolutional stem and the first two stages of ResNet-18, followed by two residual context blocks. The global diffusion
denoising network incorporates two pose-to-map attention
blocks (Vaswani et al. Vaswani, A. ; Shazeer, N. ; Shazeer, N. ; Parmar, N. ; Parmar, N. ; Uszkoreit, J. ; Uszkoreit, J. ; Jones, L. ; Jones, L. ;
Gomez, A. ;
Gomez, A. ; Kaiser, L. ; Kaiser, L. ; and Polosukhin, I. ; and Polosukhin, I. Attention
Is All You Need.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Walch, F.; Hazirbas, C. p. 8 Attention
Is All You Need. ; Hazirbas, C. ; Hazirbas, C. ; Leal-Taixe, L. 2024) has
emerged as a promising and scalable paradigm for indoor
localization by aligning egocentric RGB observations with
lightweight 2D floorplans, bypassing the need for storageintensive 3D reconstructions (Liu, Li, and Dai 2017; Sarlin
et al. 2019; Sattler, Leibe, and Kobbelt 2017) or large-scale
image databases (Balntas, Li, and Prisacariu 2018; Arandjelović et al. ; Sattler, T. ; Hilsenbeck, S. ; Hilsenbeck, S. ; and Cremers, D. ; and Cremers, D. Image-based localization
using lstms for structured feature correlation. Image-based localization
using lstms for structured feature correlation. In Proceedings
of the IEEE international conference on computer vision,
627–637.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Wüest, M.; Engelmann, F. p. 8 In Proceedings
of the IEEE international conference on computer vision,
627–637. ; Engelmann, F. ; Engelmann, F. ; Miksik, O. ; Vogel, C. ; and Pollefeys, M. ; Pollefeys, M. ; and
Barath, D. ; and
Barath, D. UnLoc: Leveraging Depth Uncertainties
for Floorplan Localization. UnLoc: Leveraging Depth Uncertainties
for Floorplan Localization. In International Conference on
Learning Representations (ICLR).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Yang, L.; Kang, B.; Huang, Z. p. 0 Compared with dense 3D models, floorplans are lightweight, easy to maintain, and readily available
in built environments, making visual FLoc particularly appealing for autonomous robot navigation (Li et al. 2025;
Huang et al. ; Huang, Z. ; and
Zhao, H. Our framework is implemented
in PyTorch and trained on a single NVIDIA RTX 5880 Ada
GPU. The visual observation encoder utilizes the frozen ViTS backbone from Depth Anything V2 (Yang et al. Depth Anything V2. ArXiv, abs/2406.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Ye, K.; Chen, M.; Zhang, X. p. 8 ArXiv, abs/2406. ; Zhang, X. From Uncertainty to Determinism: Coarse-to-Fine Visual Floorplan Localization
without Ray Matching
Shiyong Meng, Bolei Chen∗
, Ping Zhong∗
, Yang Wan, Rongzhi Wang, Jiazhi Xia, Jianxin Wang
School of Computer Science and Engineering, Central South University
{xiaowugui1017, boleichen, ping. zhong, 8205230820, 7803240121, xiajiazhi}@csu. ; and Yang, X. Fusion
of Depth and Semantics for Probabilistic Floorplan Localization. Fusion
of Depth and Semantics for Probabilistic Floorplan Localization. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 19413–
19422.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Zheng, J.; Zhang, J. p. 8 ArXiv, abs/2406. ; Zhang, X. ; Zhang, J. ; and Zhou,
Z. ; and Zhou,
Z. Structured3d: A large photo-realistic dataset for
structured 3d modeling. Structured3d: A large photo-realistic dataset for
structured 3d modeling. In Computer Vision–ECCV 2020:
16th European Conference, Glasgow, UK, August 23–28,
2020, Proceedings, Part IX 16, 519–535.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Direct pose distribution estimation allows systems to directly model
the continuous conditional pose distribution p(p I, M)
instead of relying on intermediate ray representations, which bypasses theinformation bottleneck prior to localization.
-
The coarse-to-fine architecture enables robust global hypothesis generation by using an image conditioned pose diffusion model to parameterize the continuous multimodal pose distribution across the entire floorplan.
-
The local pose refiner performs precise sub-meter correction by predicting
bounded sub-meter pose residuals from candidate-centered floorplan crops,
eliminatingstructural ambiguities
locally. -
The framework achieves real-time execution and superior localization accuracy without requiring
offline map preprocessing or test-time lookup tables,
unlike ray matching methods.
Abstract
Visual Floorplan Localization (FLoc) has emerged as a promising solution for indoor localization by matching egocentric images against minimalist structural maps. However, due to cross-modal information asymmetry and repetitive indoor layouts, visual FLoc is fundamentally challenged by multimodal pose distributions, where visually identical observations map to distinct, spatially separated locations. Existing ray-matching-based methods tackle this by explicitly predicting sparse geometric or semantic rays, which inherently incur information loss and demand resource-intensive preprocessing alongside exhaustive matching during inference. In this paper, we bypass the intermediate ray-matching paradigm and propose a coarse-to-fine visual FLoc framework that progresses from uncertainty to determinism. In the coarse stage, we design an image-conditioned pose diffusion model to parameterize the continuous multimodal pose distribution, effectively routing stochastically initialized pose particles toward distinct candidate modes. In the refinement stage, we propose a localized refiner that predicts bounded sub-meter pose residuals from candidate-centered floorplan crops, where structural ambiguities are largely eliminated. Our method effectively balances global multi-hypothesis tracking and local sub-meter refinement without requiring any offline map preprocessing or test-time lookup tables. Comprehensive results on the S3D (full) and ZInD benchmarks demonstrate that our approach achieves state-of-the-art accuracy and robustness.
Sources
- SceneAligner: 3D-Grounded Floorplan Localization in the Wild
- Floor Plan-Guided Visual Navigation Incorporating Depth and Directional Cues
- Depth Anything V2
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation