acceptodds
Under review as a conference paper at ICLR 2027

Scale-Consistent Feature Fields for Retrieval-Free Visual Relocalization

Abstract

Feature-based scene representation plays a critical role in visual relocalization, enabling robots and augmented reality systems to estimate their positions and orientations without global positioning. However, this task is often undermined by significant variations in viewpoint-induced scale and scene appearance. Recently, Gaussian splatting has emerged as a promising scene model due to its explicit geometry and real-time rasterization. In this paper, we propose ScalePlace, a retrieval-free framework that learns Scale-Consistent Gaussian Feature Fields (SCGFF) for retrieval-free centimeter-level visual relocalization. Specifically, our framework performs pose estimation through an integrated sparse-to-dense pipeline built upon a rasterizable Gaussian field, comprising three key aspects: First, Descriptor-Attached Gaussian Representation (DAGR) yields a view-dependent feature field that can be both sparsely sampled for initialization and densely rendered for refinement; Second, Multi-Resolution Feature Supervision (MRFS) facilitates cross-scale consistency through Gaussian rendering at multiple resolutions with aligned photometric and SuperPoint objectives; Third, Scale-Aware Pose Estimation (SAPE) empowers the sparse matcher and dense alignment with resolution-consistent primitives to support both initialization and iterative refinement without retrieval or pose priors. Extensive evaluations across five public datasets demonstrate that the proposed approach substantially outperforms existing baselines and ablation variants in diverse indoor and outdoor environments.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Visual Relocalization, Multi-Resolution Feature Supervision, Gaussian Splatting, Scale-Aware Pose Estimation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_V76JOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ScalePlace, a visual relocalization framework that builds upon the Gaussian feature field-based sparse initialization and dense refinement pipeline introduced by STDLoc (Huang et al., 2025). Its main contribution is a multi-resolution supervision strategy for a shared Gaussian feature field, incorporating RGB and SuperPoint feature targets at multiple rendering resolutions to encourage scale-consistent scene descriptors.

The method achieves average translation/rotation errors of 0.63 cm/0.19° on 7-Scenes and 8 cm/0.12° on Cambridge Landmarks, outperforming STDLoc on both benchmarks. Ablation studies on Chess and King’s College further show approximately 13% lower translation errors with three-scale supervision compared to native-resolution training alone, providing empirical support for the proposed multi-resolution strategy.

Critical strengths
  1. Multi-resolution supervision. The method extends STDLoc’s single-scale training by supervising a shared Gaussian feature field at multiple rendering resolutions, providing a straightforward approach to encouraging cross-scale feature consistency.
  2. Strong localization accuracy. The reported results show consistent improvements over STDLoc on both benchmarks, achieving the lowest errors across all 7-Scenes scenes and the best average accuracy among the compared NeRF/GS-based methods on Cambridge Landmarks.
  3. Detailed ablation studies. The ablation studies provide useful insights into the effectiveness of the proposed supervision strategy, with three-scale training achieving the lowest localization errors on both Chess and King’s College.
Critical weaknesses
  1. The novelty of the proposed method seems somewhat limited. Several key components, including matching-oriented landmark sampling, coarse-to-fine refinement, and localization without retrieval or pose priors, have already been explored in STDLoc. The main difference appears to be the multi-resolution feature-field supervision (MRFS), and the authors should clarify the contribution accordingly.

  2. The motivation for multi-scale supervision is not sufficiently supported by the experiments. While the ablation results show accuracy improvements, it remains unclear how single-scale training affects localization under scale changes. It would be helpful to analyze feature consistency, matching inlier rates, or localization errors at different scales.

  3. The scale robustness of the proposed method needs further investigation. Since the supervision targets are obtained by resizing a single SuperPoint feature map, it is unclear whether the improvements can generalize to scale changes caused by different camera viewpoints. Comparisons with standard image-resizing augmentation or controlled scale-change experiments would be useful.

  4. The ablation studies on training scales are only conducted on Chess and King's College. It would be better to include more scenes to verify whether the improvements are consistent across different datasets. Repeated experiments could also help assess the reliability of the reported gains.

  5. Given the similarity between the proposed pipeline and STDLoc, it is difficult to isolate the contribution of MRFS from the current comparisons. A more direct experiment would be to introduce MRFS into the original STDLoc pipeline while keeping the other settings unchanged.

  6. The additional training cost of MRFS is not discussed. Since the method requires rendering and supervision at multiple resolutions, it would be useful to report the training time and GPU memory usage. A comparison with a single-scale baseline under a similar training budget would also help assess the trade-off between accuracy and computational cost.

Questions for authors
  1. Could the authors quantify the gain from adding MRFS alone to STDLoc under otherwise identical settings? This would help isolate the contribution of the proposed supervision strategy.
  2. Does MRFS remain beneficial under controlled changes in query scale or camera distance, and how does it compare with standard image-resizing augmentation?
  3. Could the authors report the mapping time, final map storage size, and per-query localization latency, preferably with a direct comparison to STDLoc under the same hardware and evaluation settings?
  4. It would be helpful if the authors could use author–year citations consistently throughout the manuscript, distinguishing between narrative citations (e.g., “Smith et al. (2024) propose …”) and parenthetical citations (e.g., “… (Smith et al., 2024)”).
Reviewer_24p9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ScalePlace, a visual relocalization framework based on a Gaussian feature field that requires neither image retrieval nor an external pose prior. Learnable localization descriptors are attached to 3D Gaussians, allowing the same representation to support sparse landmark matching and dense feature rendering. The main identifiable addition is Multi-Resolution Feature Supervision (MRFS), which supervises the Gaussian field at multiple rasterization resolutions using RGB images and resampled SuperPoint features. At inference, landmark matching and PnP provide an initial pose, followed by iterative refinement using rendered features and depth. The paper reports average localization errors of 0.63 cm/0.19° on 7-Scenes and 8 cm/0.12° on Cambridge Landmarks. Component and training-scale ablations demonstrate improvements from the complete configuration.

Critical strengths
  1. Competitive reported localization accuracy. In Table 1, ScalePlace improves both translation and rotation errors over STDLoc across all seven indoor scenes. Table 2 also reports competitive average outdoor accuracy. These improvements would have practical value.
  2. The additional supervision does not require extra inference rendering scales. MRFS supervises a shared Gaussian representation at multiple resolutions during mapping. Table 4 shows that three-scale supervision outperforms the evaluated single-scale and two-scale configurations, providing initial evidence for the usefulness of this training strategy.
Critical weaknesses
  1. The distinction from STDLoc is insufficiently clear, and the incremental contributions need better justification. STDLoc already includes a Gaussian feature representation, matching-oriented landmark sampling, a scene-specific detector, and coarse-to-fine pose refinement using rendered features and depth. These components substantially overlap with the main operations of DAGR, MOS, and SAPE. Therefore, presenting matching-oriented sampling and coarse-to-fine refinement as additions relative to STDLoc is insufficiently justified. The authors should explicitly distinguish inherited components from substantive algorithmic changes. Based on the current description, MRFS is the main clearly identifiable addition, while the presentation risks overstating the scope of the methodological novelty.
  2. The experimental validation of MRFS, the main new component, remains incomplete. The ablations show that adding MRFS reduces localization errors, which is encouraging. However, MRFS introduces both image and feature supervision at multiple resolutions, without separately evaluating their contributions. It therefore remains unclear whether the improvement primarily comes from better image and geometry reconstruction or from the multi-resolution feature learning emphasized by the paper. Separate controls using only additional multi-resolution image supervision and only additional multi-resolution feature supervision would clarify this issue. Furthermore, the evaluation mainly reports aggregate localization accuracy without separately examining queries with large scale changes, leaving the central claim of scale robustness without direct experimental validation.
Questions for authors
  1. Please clarify the specific algorithmic differences between DAGR, MOS, and SAPE and their counterparts in STDLoc. In particular, what substantive new operations are introduced in MOS’s scoring and neighborhood selection, and in SAPE’s coarse-to-fine matching and pose updates?
  2. Are the STDLoc results in Table 2 taken from the original paper or reproduced? Do both methods use the same refinement counts, image resolutions, preprocessing, and computational budgets? What improvement does MRFS provide when the refinement count is matched?
  3. Could you provide localization results grouped by scale or viewpoint differences between query and training views, together with multi-resolution RGB-only, feature-only, and compute-matched single-resolution training controls? These results would materially affect my assessment of the core contribution.
  4. In the “w/o DAGR” configuration in Table 3, what representation and descriptors replace the original Gaussian feature field? How do MOS and SAPE operate under this configuration?
Reviewer_ZjbROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a scene specific visual relocalisation method built on feature Gaussian splatting based map. Each Gaussian carries a descriptor that is supervised by a frozen teacher (SuperPoint). The main new idea is to render the field at 3 different resolutions during mapping and supervise each rendering with photometric and feature targets, which they call it multi-resolution feature supervision.

After mapping, landmarks are sampled using a matching score, a detector is trained to find the landmarks in query images, an the pose is initialised by a PnP solver and refined via a few coarse to fine matching against rendered feature and depth maps. There is no prior pose and experiments on two datasets show small improvements over the baseline (STDLoc), as the closest prior work.

Critical strengths

The method description is clear enough and shares enough implementation details and the protocols used. The experiments and comparisons are okay and cover a few various baselines including NeRF/GS and the reported results are the best in the NeRF/GS category on both datasets. The ablations on training scales and sampling methods are also informative.,

Critical weaknesses

I actually have some major issues about the novelty over the closest prior work STDloc and how it is presented. Stdloc already introduces a feature Gaussian representation, matching oriented sampling, scene specific detector and also sparse to dense pipeline with feature matching, also without pose prior, the same components presented in the ablation study with different labels such as DAGR, MRF, ... so I'm not sure how they are different from the contributions of STDloc?

If I understood correctly, the actual contribution is multi-resolution feature supervision, which the ablations show it is the module with the smallest impact. the large drop in the performance comes from removing the parts that stdloc already has.

The other issue is about scale consistency. I don;t see any support for that claim. the field is rendered at a lower resolution and compared to bilinearly resized features, there is no experiment that isolates scale change for example queries at a different resolution that the mapping views.

The gains are small and no variance over runs is reported. for example On 7-Scenes the differences to stdloc are fractions of a millimeter, which is likely below the accuracy of the ground truth itself, so it is hard to say whether these are real improvements.

Questions for authors
  • Please clarify exactly what your method adds over stdloc?
  • Did you try supervising with SuperPoint run on the resized images or any other way of injecting real scale change?
  • Can you add an experiment that directly tests scale robustness? e.g queries taken at a different resolution or distance from the mapping views with and without your multi resolution supervision?
  • In the ablation study, when you remove the dagr and sape, what replaces the removed part?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.