acceptodds
Under review as a conference paper at ICLR 2027

RAGE-I2P: Reliability-Aware Geometric Embeddings for Image-to-Point Cloud Registration in Changing Environments

Abstract

Existing image-to-point cloud registration methods have achieved strong performance in static scenes. However, real-world environments often contain moved, removed, or newly added objects. Such changes can preserve feature similarity while breaking physical correspondence, producing unreliable matches at both coarse and fine levels. We propose RAGE-I2P, a geometry-guided framework for reliable registration in changing environments. We learn coarse- and fine-level geometric embeddings from image and point-cloud normals. At the coarse level, a covariance-aware discrepancy between geometric embeddings provides compatibility weights for cross-modal attention. At the fine level, normal-derived embeddings enrich feature extraction with local surface information, while an additional embedding loss encourages correspondence-discriminative representations. We also propose ASIMLab, a seven-scene dataset captured under real scene changes, and construct Changing RGBDScenesV2 through controlled image edits. Experiments on both benchmarks show that RAGE-I2P achieves the best mean correspondence and registration performance. The source code and the datasets used in the paper will be released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image-to-Point Cloud Registration, Changing Environment, Geometric Embeddings

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6rDTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAGE-I2P for image-to-point cloud registration in changing environments. It uses normal-derived geometric embeddings to improve coarse- and fine-level correspondence reliability, and introduces two benchmarks for evaluation. Experiments show consistent improvements over recent baselines in correspondence quality and registration recall.

Critical strengths
  1. The paper addresses a practically relevant problem. Image-to-point cloud registration under changing environments is important for long-term localization and robotic perception. The paper clearly motivates why appearance similarity alone can become unreliable when objects are moved, removed, or newly introduced.
  2. The experimental evaluation is reasonably comprehensive. The method is evaluated on both a synthetically modified benchmark and a newly collected real-world dataset, with ablation studies and efficiency analysis. The proposed method also shows consistent improvements over several recent baselines in correspondence quality and registration recall.
Critical weaknesses
  1. The methodological novelty appears relatively limited. The proposed framework mainly extends an existing coarse-to-fine I2P registration pipeline by introducing normal-derived geometric embeddings, geometric compatibility weighting, and an additional embedding loss. Although these components are reasonable, the overall technical advance appears incremental rather than fundamentally different from existing geometry-aware correspondence learning approaches. This is the main factor limiting my overall assessment.
  2. The evidence for generalization to diverse real-world changing environments is still limited. The real-world ASIMLab benchmark contains only seven scenes and is restricted to tabletop regions, while Changing RGBDScenesV2 is constructed through synthetic image editing. Therefore, although the reported improvements are encouraging, it remains unclear how robustly the method generalizes to more diverse and complex real-world scene changes.
Questions for authors
  1. Clarification on methodological novelty: Could the authors clarify more explicitly what the key technical distinction is between RAGE-I2P and existing geometry-aware or normal-guided correspondence learning methods? In particular, which component should be regarded as the main methodological contribution beyond incorporating normal-derived geometric information into a coarse-to-fine registration framework?
  2. Clarification on the source of performance gains: Table 3 shows that adding DSINE normals to 2D3D-MATR already improves RR substantially from 25.8% to 44.2%, while the full RAGE-I2P further improves it to 49.8%. Could the authors clarify how much of the overall improvement should be attributed to the proposed reliability-aware geometric modeling rather than to the strong pretrained normal prior itself?
Reviewer_1PQEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies image-to-point-cloud registration after scene changes. It builds coarse and fine geometric embeddings from image and point-cloud normals. Coarse embedding discrepancies reweight cross-attention, and fine embeddings are concatenated into the features and supervised by an embedding loss. On edited RGBDScenesV2 and the collected ASIMLab dataset, the method reports the best mean results among the compared baselines.

Critical strengths
  1. The task is meaningful. Existing I2P benchmarks are captured at the same time without scene changes, whereas ASIMLab includes object addition, removal, and movement.
Critical weaknesses
  1. The writing is vague and insufficiently rigorous, and it contains errors. The notation in the method section does not match Figure 2, and writing FF as 5-dimensional contradicts the high-dimensional embedding GG.
  2. The method does not adequately solve registration in dynamic scenes. It is not evaluated on outdoor dynamic data, which would be a more suitable setting. The reported reliability comes mainly from the frozen normal estimator: the proposed modules add little, and the evaluation protocol is too weak to support the claim. In dynamic scenes, feature similarity between corresponding objects is the primary cue, while normals are only auxiliary.
Questions for authors
  1. Please align the symbols in Figure 2 with the equations, and clarify the dimension of GG and how FF can be 5-dimensional.
  2. How are correspondences on moved objects handled, and how are they prevented from corrupting the RT estimate?
  3. Why is the method not evaluated on outdoor datasets that contain dynamic objects, such as KITTI?
Reviewer_8QqPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes RAGE-I2P for image-to-point-cloud registration in dynamic environments. The method introduces geometric embeddings derived from surface normals, structured into coarse-grained and fine-grained levels. Furthermore, the paper introduces ASIMLab—a dataset comprising seven scenes—and presents an enhanced version of the RGBDScenesV2 benchmark, constructed through image-editing techniques. Experimental results indicate that the proposed method outperforms existing state-of-the-art approaches.

Critical strengths

The problem setting is practically significant, and the experimental results are promising.

Critical weaknesses

1. The methodological innovation is not sufficiently justified.

LDF-I2P already incorporates surface normals from both images and point clouds into framework. Consequently, the use of geometric information derived from normals is not an innovative contribution of RAGE-I2P. While some enhancements are logically sound, the paper fails to adequately demonstrate that they offer a significant advantage over simpler extensions of LDF-I2P.

2. Insufficient validation of robustness against real-world scene changes.

The "Changing RGBDScenesV2" dataset is constructed by editing images while retaining the original point clouds and camera parameters, with depth maps estimated from the edited images. While this controlled experimental setup has its merits, the geometric inconsistencies and artifacts introduced by image editing and depth estimation may differ from the scene changes that occur naturally in the real world. The ASIMLab dataset contains only seven scenes—one used for training and the rest for evaluation. Consequently, there is insufficient evidence to support the method's ability to generalize to diverse, real-world environments.

Questions for authors
  1. Given that LDF-I2P already leverages surface normals via the GAFE module, what specific advantages does the proposed "learned normal embedding" offer over direct fusion of normal features? In the existing ablation studies, which experiment specifically isolates the contribution of the "covariance-aware compatibility weighting"?

  2. How do the authors ensure that the results on the "Changing RGBDScenesV2" dataset are not significantly affected by the inconsistencies between the geometry of the edited images and the estimated depth maps?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.