acceptodds
Under review as a conference paper at ICLR 2027

GCD-SLAM: Geometric-Correction Decoupling for RGB-D Gaussian Splatting SLAM

Abstract

Localization refinement in evolving scene representations can become inherently coupled with the representations used for optimization, when these representations are constructed from previously estimated states. Such coupling makes localization correction dependent on the underlying representation and its construction history, potentially limiting the effectiveness of state refinement. We identify this representation-dependent optimization issue in RGB-D Gaussian Splatting SLAM and propose **GCD-SLAM**, a Geometric-Correction Decoupling framework that introduces representation-independent localization correction beyond the native Gaussian tracking-mapping loop. Specifically, GCD-SLAM develops **Independent Geometric Guidance** to derive correction cues directly from temporal depth observations rather than trajectory-conditioned Gaussian representations. **Reliability-aware Correction Selection** identifies reliable correction conditions and enables selective localization refinement. Finally, **Localization State Separation** decouples corrected poses from the native tracking-mapping process, preserving Gaussian reconstruction consistency while providing corrected localization outputs. Experiments on challenging RGB-D SLAM scenarios demonstrate that GCD-SLAM achieves effective selective localization improvement while preserving the native Gaussian tracking-mapping process and maintaining robustness under challenging near-range conditions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D Gaussian Splatting, RGB-D SLAM, camera localization, geometric correction, tracking-mapping decoupling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_hX36OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GCD-SLAM, which corrects the output pose of Gaussian SLAM through independent depth geometric constraints, reliability screening, and separation of positioning states. The problem is of certain significance, but the current experimental results and mechanism analysis are not yet sufficient to support the main contribution of the paper.

Critical strengths

None

Critical weaknesses

The experimental coverage and actual benefits are rather limited.

There is a lack of sufficient comparison under a unified protocol.

The necessity of the reliability screening module has not been fully supported.

There are also no directly comparable rendering results.

Questions for authors
  1. Why did the authors only provide a small number of results for the TUM RGB-D dataset and did not conduct quantitative analysis of rendering quality? The existing experiments do not seem to support the authors' claims.
  2. Table 1 mainly compares the original VarSplat with the results after adding corrections. Other SLAM methods are placed in the appendix, and the running and statistical protocols are different. The current results make it difficult to judge the relative competitiveness of the method compared to the existing system, and there is also a lack of verification on other Gaussian SLAM backbones.
  3. The necessity of the reliability screening module has not been fully supported. In Table 2, the improvement without using RCS is 34.961%, and with RCS it is 34.810%. This does not reflect the performance advantage of this module.
Reviewer_6he9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GCD-SLAM examines whether a Gaussian map built from earlier pose estimates can influence which new pose its tracking objective prefers. The authors compare maps built from estimated and ground-truth poses while holding the observations and candidate poses fixed, then add a separate correction to a VarSplat RGB-D system. That correction uses consecutive depth frames to estimate a small geometric adjustment, admits it under fixed texture, range, and support rules, and applies it only to the reported trajectory; native tracking and mapping continue with the original poses. Across three seeds, median ATE on a selected 300-frame near-range segment falls from 1.743 to 1.137 cm. Over the full 1,060 frames, the median relative reduction is 0.414%, and the three standard public sequences in Table 1 do not activate the correction.

Critical strengths
  1. The fixed-probe experiment holds more constant than an ordinary trajectory comparison. All 12 tested pose pairs change preference when the Gaussian reference changes. I read this as useful evidence that the tested objective depends on how its map was constructed, while appreciating that the paper does not claim this experiment alone explains the final trajectory error.

  2. The near-range result is also more credible because the authors report all three seeds and replay the correction from recorded candidates. ATE improves on the 300-frame prefix in every seed. That is a meaningful result for the stated condition, and the paper provides enough detail to see where it stops holding.

Critical weaknesses
  1. The localization benefit is too concentrated to support a broad SLAM claim. The three public sequences in Table 1 are unchanged, the median improvement over the full near-range sequence is only 0.414%, and the room stress test regresses slightly. The near-prefix rotational RPE at one frame also worsens. With one Gaussian backbone and a selected short horizon carrying most of the gain, I cannot tell how often this correction would help in ordinary use.

  2. The experiments do not convincingly separate the proposed three-part framework from a simpler output-side depth correction. Admitting every valid candidate gives a 34.961% near-prefix ATE reduction, compared with 34.810% using reliability selection. Point-to-plane depth registration is an established RGB-D tracking cue, including in RGBD GS-ICP SLAM; the paper’s Open3D ICP control changes both solver and policy, so it is not a matched test of the distinctive design. Finally, identical mapping scores follow from evaluating the same map and views. In the frozen-map audit, the corrected poses actually have a slightly higher depth residual.

Questions for authors
  1. Why does the correction help throughout the 300-frame prefix but yield little benefit by frame 1,060? Were the activation thresholds or the 300-frame horizon chosen after inspecting this sequence?

  2. Can you compare against a simple output-only depth registration baseline using the same geometric candidates, bounds, and evaluation protocol, both with and without reliability selection?

  3. Does the map-reference preference reversal occur on held-out observations and other scenes or backbones? Is its magnitude associated with the frames on which geometric correction helps?

Reviewer_v4qkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents an RGB-D based 3D Gaussian splatting SLAM method that mitigates Representation-Induced Localization Bias (RILB), where camera tracking becomes skewed by its own evolving map. To resolve this, this paper introduces a three-stage decoupling framework: 1) Independent Geometric Guidance extracts pose corrections directly from temporal depth images via point-to-plane ICP. 2) Reliability-aware Correction Selection filters frames using heuristic thresholds. 3) Localization State Separation outputs corrected poses without updating the native map.

Critical strengths

On challenging near-range, low-texture TUM RGB-D sequences where appearance-based Gaussian tracking struggles, the proposed geometric correction achieves a 34.81% median ATE reduction on the 300-frame prefix while keeping the native Gaussian map and 3D rendering quality unchanged.

Critical weaknesses
  • The core technical components of the proposed method are remarkably incremental and rely on established, standard operations. 1) IGG is essentially a standard Point-to-Plane ICP algorithm (with GNC-TLS weighting) applied to consecutive temporal depth frames. 2) RCS consists of simple, hand-tuned heuristic thresholds. 3) LSS / Zero Writeback merely skips updating the native tracker and map representation with the corrected pose. While the paper introduces an extensive array of newly coined terminology (e.g., RILB, IGG, RCS, LSS), the actual underlying algorithms lack technical novelty. The academic framing over-complicates what is fundamentally a combination of standard depth registration, heuristic filtering, and omitting a state feedback step.
  • What the paper introduces as "Representation-Induced Localization Bias (RILB)" is a standard, well-known issue in frame-to-model RGB-D SLAM. When a dense map is incrementally fused from estimated poses, tracking errors corrupt the map and bias subsequent pose estimates toward the drifted geometry.
  • The quantitative evaluation is severely limited and selective. The primary 34.81% improvement comes from the first 300 frames of one specific near-range sequence (fr3/structure_notexture_near) in the TUM RGB-D dataset. Over the full sequence, the gain is only 0.414%. There is no held-out validation for the heuristic thresholds, no Replica or ScanNet, and only one backbone.
Questions for authors
  • Can the authors clarify the core algorithmic novelty of IGG, RCS, and LSS beyond combining standard consecutive-frame Point-to-Plane ICP, hand-tuned heuristic thresholds, and omitting tracker/map feedback?
  • How do the authors justify the practical effectiveness and generalization of the method? Can the authors evaluate the method on full sequences of standard benchmarks such as Replica or ScanNet, and provide results on an additional 3DGS SLAM backbone beyond VarSplat to show the generalizability of the proposed components?
  • Can the authors provide held-out validation for the heuristic thresholds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.