acceptodds
Withdrawn from ICLR 2027

PanoWalker: Iterative Persistent Panoramic Scene Construction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D scene reconstruction is increasingly required in real-world mobile capture, where users progressively explore environments and casually acquire observations along a navigation trajectory. Such capture calls for a scene representation that can grow as observations arrive. Panoramic imagery is especially attractive in this regime as each observation captures the full surrounding environment. We present PanoWalker, a lightweight, training-free pipeline for iterative panoramic 3D scene reconstruction. Given a sequence of panoramic pairs, PanoWalker uses a frozen panoramic feed-forward backbone to predict one local Gaussian set for each pair and then incrementally fuses these local predictions into a persistent 3D Gaussian scene. The persistent scene retains the source identity of each pairwise prediction. Optional geometric cleanup screens incoming candidates and constrains the accumulated scene around the capture trajectory; the indoor headline setting disables these filters. At rendering time, source-distance opacity reweighting improves the evaluated near-input HM3D views without changing the stored scene. The pipeline uses frozen backbones and extends the scene without re-predicting earlier pairs or performing per-scene optimization. It leads on the HM3D and Replica aggregate benchmarks, while outdoor final-quality results are mixed. In a separate progressive-preview task on two 360Loc segments, it delivers three successive scenes from 10, 20, and 30 panoramas in 126 s on average. Under approximately the same cumulative time budget, its final WS-PSNR is 17.61 dB, compared with 14.42/13.67 dB for continued ODGS/360GS optimization. The code and model will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Panorama, Gaussian Splatting, 3D reconstruction

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_qruEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PanoWalker, a training-free framework for incremental panoramic scene construction. The method uses a frozen pairwise Gaussian prediction backbone to reconstruct local Gaussian sets from panoramic image pairs. These predictions are accumulated in a shared world coordinate while preserving their source identities. To mitigate geometric inconsistencies across predictions, the method independently renders each Gaussian set and combines the results using distance/coverage-aware blending. Experimental results show that PanoWalker outperforms existing baselines on most benchmarks. Moreover, the method enables substantially faster progressive scene updates than optimization-based approaches while maintaining rendering quality.

Critical strengths
  • Practical problem setting: The paper addresses incremental panoramic scene reconstruction, a practical yet relatively underexplored setting. The proposed framework enables progressive scene construction without requiring per-scene optimization.

  • Simple and effective framework: PanoWalker provides a straightforward way to extend existing feed-forward Gaussian prediction models to persistent scene reconstruction. Its training-free design also makes the framework easy to integrate with pretrained backbones.

  • Effective source-aware rendering: The proposed rendering strategy addresses geometric inconsistencies between independently predicted Gaussian sets. Independent rasterization and distance/coverage-aware blending improve rendering quality without modifying the accumulated scene.

Critical weaknesses
  • Limited methodological novelty: The proposed framework combines existing pairwise Gaussian prediction with relatively simple rendering and blending strategies. While effective, the technical novelty of the individual components appears somewhat limited.

  • Unclear comparison setting: In Table 1, PanoWalker uses more input panoramas than some of the compared feed-forward methods. While 360GS and ODGS use a similar number of input panoramas to PanoWalker, other feed-forward baselines, such as PanSplat, MVSplat360, and Splatter-360, use fewer input views. In particular, PanoWalker adopts Splatter-360 as its backbone, but a comparison with Splatter-360 using the same 30 input views is missing. This makes it difficult to determine whether the performance gains come from the proposed framework or the larger number of input observations.

  • Insufficient ablation studies: The ablation study focuses only on distance and coverage weighting, despite the framework consisting of several components. More comprehensive and controlled ablation studies would help better justify the effectiveness of each component.

Questions for authors
  • Robustness analysis under camera pose error: Although known camera poses are also assumed by the compared baselines, PanoWalker additionally relies on globally consistent poses for persistent Gaussian accumulation and distance-based blending. In casual walk-and-build mobile capture scenarios, camera poses are often subject to estimation errors and accumulated drift. It would be valuable to evaluate the robustness of the proposed framework under noisy camera poses to better demonstrate its applicability to real-world scenarios.
Reviewer_HD1sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PanoWalker, a training-free framework for progressive panoramic scene reconstruction. Given sequentially acquired panoramas with known camera poses, it incrementally accumulates Gaussian predictions from a frozen pairwise backbone into a persistent scene. At query time, independently rendered source layers are combined using camera-distance and opacity-coverage weights, avoiding per-scene optimization or modification of earlier predictions. Experiments demonstrate strong novel-view synthesis performance across four benchmarks and fast incremental updates on two walking trajectories.

Critical strengths

1. Meaningful Progressive Reconstruction Setting. The paper addresses a practical setting largely overlooked by conventional one-shot reconstruction methods: maintaining and extending a reusable scene as new observations arrive.

2. Strong System-Level Performance. The approach achieves substantial rendering improvements over the evaluated baselines without additional training or per-scene optimization. Its ability to reuse previous predictions enables fast incremental updates, demonstrating the practical potential of combining pretrained reconstruction models with persistent scene accumulation.

Critical weaknesses

1. Limited Algorithmic Novelty. The proposed method does not introduce a new mechanism for learning or constructing a unified 3D scene.

  • Its persistent representation simply concatenates pretrained pairwise Gaussian predictions, while cross-source inconsistencies are handled through query-time image blending rather than geometric fusion.

  • The central composition strategy-weighting independently rendered layers by camera distance and coverage-closely follows established local image-based rendering techniques such as LLFF.

The contribution appears to be primarily a system-level adaptation of existing components, rather than a substantive advance in 3D reconstruction or scene representation for ICLR.

2. The Gains Are Not Convincing. The reported improvements are substantial, but the experimental evidence does not establish that they arise from the proposed composition strategy.

  • Confounded input comparison: On HM3D, the reported +7.27 dB PSNR gain over multi-view Splatter-360 compares approximately 30 distinct input panoramas against only 15 context views.

  • Modest contribution from the proposed blending: With an identical Gaussian bank, Table A4 shows only 0.18–0.37 dB improvement over straightforward pairwise image overlay across the four benchmarks.

These results suggest that the impressive overall performance may be driven largely by accumulated observations and proximity-based source selection, rather than a genuinely new reconstruction mechanism. The claimed methodological advantage remains insufficiently demonstrated.

Questions for authors

Q1. Methodological novelty. Beyond concatenating pretrained Gaussian predictions, what is the essential technical contribution that distinguishes PanoWalker from existing local image-based rendering approaches, particularly LLFF-style distance/coverage-weighted blending?

Q2. Performance attribution. Can the authors provide input-matched comparisons using identical panoramas, backbone predictions, and scoring resolutions? Given the relatively small improvement over simple layer overlay in Table A4, what evidence supports attributing the headline gains to the proposed composition strategy?

Q3. Genuine novel-view generalization. How does performance change when excluding query cameras co-located with input cameras? Can the authors report results on independently selected off-path queries at increasing distances from the observed trajectory?

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

The paper presents a training-free framework for incrementally constructing panoramic Gaussian scenes by accumulating pairwise predictions and blending independently rendered sources. However, near-input evaluation, unequal input budgets, scalability, and novelty over existing blending methods require further clarification. If the authors can address these concerns with additional experiments and analysis, I would be willing to reconsider my rating.

Critical strengths
  • The paper addresses a practical scenario in which users progressively capture panoramas and expect an expanding, reusable scene. Accumulating predictions from frozen pairwise backbones avoids additional training and repeated scene optimization. Preserving source identities and combining independently rendered layers according to camera distance and coverage provides a coherent approach to handling overlapping predictions while retaining previously reconstructed content.
  • The experimental results are promising across several benchmarks. PanoWalker leads the reported metrics on HM3D, Replica, and 360Loc, and achieves the lowest LPIPS on Insta360. The progressive evaluation also goes beyond conventional final-scene testing by measuring successive updates and previews. Three updates take 56.6 seconds in total on average, demonstrating the practical benefit of reusing earlier predictions.
Critical weaknesses
  • The novel-view evaluation is concentrated near input viewpoints. On HM3D, the median nearest-input distance is only 0.020 m, and 26 of 80 query centers coincide with an input center. Off-path queries were used during method development, limiting independent evidence of performance farther from the capture trajectory.
  • The main comparison uses unequal input budgets: PanoWalker processes approximately 29–30 panoramas, whereas multi-view Splatter-360 uses 15. Moreover, jointly rendering the same Gaussian bank already achieves 30.33 dB versus 30.53 dB for PanoWalker on HM3D, leaving the additional contribution of the proposed composition rule insufficiently established.
  • Scalability remains a concern. With 30 inputs, the method stores 53.19 million Gaussians and uses approximately 23 GB of CPU memory. Every query processes all source pairs, and the final-stage first-preview component takes 4.67–5.30 seconds. Longer trajectories could therefore incur substantial storage and rendering costs.
  • The methodological novelty requires clearer positioning. Independent local rendering with camera-distance and accumulated-opacity weighting has precedents, including LLFF, as acknowledged in the paper. Stronger comparisons with simple source-selection and blending baselines would help establish the necessity of the proposed design.
Questions for authors
  • Can the authors evaluate on an independent set of queries farther from the input cameras and capture trajectory? Please report results grouped by nearest-input distance.
  • Can the authors compare methods using the same unique input panoramas?
  • How do CPU memory, GPU memory, update latency, and rendering latency scale with trajectory length and input count?
  • What distinguishes the proposed rendering strategy from existing distance- and opacity-weighted blending approaches such as LLFF?

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

Related papers

Discussion (0)

Sign in to comment.