acceptodds
Withdrawn from ICLR 2027

Incremental Reconstruction of Long-Range Panoramas with Global Consistency

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Long-sequence panoramic reconstruction requires detailed scene geometry and globally consistent camera poses across distant revisits. Panoramic SLAM and incremental SfM rely on feature correspondences, but repeated, similar structures and small partial overlap make reliable long-range connections difficult to establish, leaving accumulated drift unresolved. Feed-forward models offer detailed local geometry, while their high computational costs bounded the inference windows, which require additional constraints for global consistency. This motivates combining continuous feature-based tracking and sparse mapping with rapid, high-quality local reconstruction from feed-forward models to achieve accurate localization and globally consistent dense mapping over long sequences. We therefore propose a panoramic reconstruction framework that couples a ORB-based sparse panoramic mapping frontend with a fine-tuned feed-forward model. Our key insight is that predicted local geometry enables 3D verification of feature matches across distant revisits, yielding long-range camera constraints even between regions. We calibrate predicted depths with sparse landmarks and use verified connections to optimize the global pose graph, updating dense geometry with the corrected camera poses. We also introduce PanoRecon, a seven-scene benchmark with repetitive architecture, large and small loops, and reference camera poses and LiDAR geometry for evaluating global consistency. Across 360LOC, YUTO MMS, and our proposed PanoRecon benchmarks, our method achieves the lowest mean absolute trajectory error among evaluated methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Panorama, 3D Reconstruction, 360 Images, Large-Scale Reconstruction

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

This paper presents an incremental panoramic reconstruction framework that combines an ORB-based panoramic SLAM frontend with Pi3-based dense geometry prediction. Sparse landmarks calibrate the predicted depths, while predicted geometry supports the verification of long-range connections for global refinement. The paper also introduces PanoRecon, a benchmark for evaluating reconstruction consistency over long panoramic sequences.

Critical strengths
  1. The paper presents an extensive experimental evaluation.
  2. The proposed PanoRecon dataset is meaningful and provides a useful benchmark for long-range panoramic reconstruction and global consistency.
Critical weaknesses
  1. The methodological novelty is limited. The overall framework combines an ORB-based panoramic SLAM frontend for pose estimation with Pi3 for dense reconstruction. Related combinations of sparse geometric tracking and feed-forward reconstruction already exist, including EC3R-SLAM [4], SwiftMap [5], and SLAM-MER [6], although these primarily use conventional perspective images. Moreover, the proposed dense reconstruction pipeline converts panoramas into perspective views before applying the existing reconstruction model. The paper therefore needs to establish more clearly what methodological advances arise specifically from handling panoramic inputs.

    The main methodological contribution appears to lie in integrating these components. However, several key parts of this integration have close precedents:

    a) Partitioning long sequences into local submaps, reconstructing them separately, and subsequently aligning and optimizing them: VGGT-SLAM [7] and EC3R-SLAM [4].

    b) Using SLAM/VIO geometry to calibrate predicted depth on a per-frame or per-keyframe basis: HI-SLAM [8], DeepRelativeFusion [9], and VI-Depth [10].

    c) Using voxel overlap for submap association: related mechanisms have been used for historical submap association [2], loop verification in KISS-SLAM [1], and submap fusion [3]. Although their precise roles differ, the underlying overlap-based association idea has clear precedents.

    These precedents establish that several constituent design choices are not new. Overall, the paper does not sufficiently distinguish its methodological innovation from the panoramic adaptation and integration of existing techniques.

  2. Graph correction, landmark observation fusion, and bundle adjustment are established components of SLAM systems. Their use here does not, by itself, establish substantial methodological novelty.

References

[1] Guadagnino, T., et al. "KISS-SLAM: A Simple, Robust, and Accurate 3D LiDAR SLAM System With Enhanced Generalization Capabilities." IROS, 2025.

[2] Koide, K., et al. "Globally Consistent 3D LiDAR Mapping with GPU-accelerated GICP Matching Cost Factors." IEEE RA-L, 2021.

[3] Wang, Y., et al. "Scalable and Elastic LiDAR Reconstruction in Complex Environments Through Spatial Analysis." ECMR, 2021.

[4] Hu, L., et al. "EC3R-SLAM: Efficient and Consistent Monocular Dense SLAM with Feed-Forward 3D Reconstruction." arXiv:2510.02080, 2025.

[5] Xu, J., et al. "Towards Fast and Fully Automatic Drone Mapping." ACM MobiSys, 2026.

[6] Piedade, V., et al. "Revisiting Monocular SLAM with Spatio-Temporal Scene Modeling." CVPR, 2026.

[7] Maggio, D., et al. "VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold." NeurIPS, 2025.

[8] Zhang, W., et al. "HI-SLAM: Monocular Real-Time Dense Mapping With Hybrid Implicit Fields." IEEE RA-L, 2024.

[9] Loo, S. Y., et al. "DeepRelativeFusion: Dense Monocular SLAM using Single-Image Relative Depth Prediction." IROS, 2021.

[10] Wofk, D., et al. "Monocular Visual-Inertial Depth Estimation." ICRA, 2023.

Questions for authors
  1. Is the original SLAM frontend’s built-in loop detection module enabled in the proposed system?
Reviewer_f4FWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduce a pipeline that cooperate conventional feature based method with the 3D foundation models, and a challenging PanoRecon benchmark to evaluate the trajectory and reconstruction.

Critical strengths

The collaboration design between conventional sparse method and 3D foundation dense model is interesting for SLAM community. The challenging benchmark could futher evaluate the upper limit of existed methods.

Critical weaknesses

Major concerns:

  1. The related work does not clearly position the contribution of this work.

  2. The motivation of keeping conventional component is not strong. Comparing with 3D foundation models, feature matching naturally suffer more on textureless, motion blur and lighting changes. I donot see the motivation on using feature matching comparing with HALO-SLAM's foundation based tracking that also deal with global consistency.

  3. For the global consistency, I only seem optimization on poses (scale, translation, rotation), but no geometric refinement after the Pi3's depths.

  4. Since the testing would like to emphasize the advances on challenging SLAM scenes. The challenging benchmark should also include the commen challenges such as motion blur or lighting changes, which may make feature-based really hard.

  5. Paper does not compare PanoVGGT on existed localization benchmarks, such as Matterport3D, Stanford2D3D and PanoCity. Only shows a OOM doesnot give a quality comparison. Paper does not compare HALO-SLAM on existed localization benchamrks, such as Holo360D, PanoVILD, PAIR360, 360-VIO and 360Loc.

  6. No reconstruction comparison with Feed-Forward SOTAs, such as PanoVGGT.

Minor concerns:

  1. Speed and memory cost is not listed.
Questions for authors
  1. Clarify the motivation of using conventional feature matching. and etc. together with 3D foundation model. Addressing the global consistency concern.
  2. Comparing with SOTAs.
Reviewer_9eAKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a method for long-sequence panoramic reconstruction as well as PanoRecon, a benchmark with repetitive architecture and long-range revisits to evaluate performance on this task. The methodology combines the strengths of classical SLAM / SfM with those of feed-forward models: specifically, classical methods can establish global consistency and loop closure, but are weak to low-texture / repeated texture surfaces, whereas feed-forward models may be able to deal with these cases better but have computational limits that allow them to be run only locally.

Critical strengths

The overall intuition of the reconstruction method’s combination of classical SLAM / SfM with feed-forward methods is intuitive and seems to be the right approach. The weaknesses of each type of approach make sense, and the combination of the two allow each approach to address the weaknesses of the other.

Critical weaknesses
  • The paper is incredibly hard to read even with careful reading. Specifically, most technical details are glossed over too quickly without establishing the intuition of what is going on. For example:
    • L.142-145: the authors mention using ORB-based ERP frontend and Pi3 fine-tuned on Holo360D; however, the basic mechanisms, or at least the input / output of these methods, are all not established, so the manuscript is not self-contained. Fine-tuning setup is not discussed at all. The order of local / global optimization of the ORB method vs. Pi3 application is also not discussed; Fig. 2 does not help, as it just has two arrows going between both modules.
    • L.216-219: it seems like the authors claim that their method detects repeated structures and see if they produce actual or spurious matches (a challenge identified on L.35-36). However, the authors then present a list of concepts such as “bidirectional distance,” “nondegenerate geometry,” “distinct feature support from multiple keyframes on both sides” (sides of what?), without giving any explanation of how these are specifically implemented or used to do the detection. Appendix B.2 does not seem to establish the connection between these concepts and the spurious repeated structure detection problem; it seems to be mostly a list of technical equations and thresholds.
  • Novelty / purpose of PanoRecon is questionable. PanoRecon is supposed to stress test repetitive architecture and long-range revisits for trajectory evaluation, which SLAM / SfM methods should struggle with (L.35-37). However, the performance of the SLAM and SfM methods in Table 1 and Fig. 5 show that these methods perform comparably to the method presented by the authors, and in some cases outperform the presented method on certain metrics like RPE-t and RPE-r; so this calls into question whether PanoRecon contains challenging enough repetitive architecture / long-range scenes to induce a performance drop for SLAM / SfM. In contrast, on YUTO MMS, the proposed method does much better than these SLAM / SfM methods. Fig. 5 shows that Residence2 from PanoRecon has qualitatively similar results across the board; but Sequence C from YUTO MMS is significantly harder.
Questions for authors
  • Clarification request: can you clarify in a more intuitive manner how your proposed method works? For example, in what order are each component of ORB-based ERP frontend and Pi3 run in, how does the sky segmentation network (L.149-151) fit into the picture, how is repeated structure filtered for spurious vs. actual matches, etc?
  • Can you quantify the qualities of PanoRecon that distinguish it from previous benchmarks such as YUTO MMS and provide a comparison? Specifically, how much low-texture/repeated/similar structures is there, how limited is the overlap for loop closure? If it is hard to quantify, qualitative examples / side-by-side comparisons of the raw benchmark images could also be helpful instead.
  • Can you explain why the performance of SLAM / SfM methods seems to be very close to the performance of your proposed method on PanoRecon despite its emphasis on repeated / similar structure and small partial overlap etc?

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

Related papers

Discussion (0)

Sign in to comment.