acceptodds
Under review as a conference paper at ICLR 2027

Decoupling Spherical Reasoning from Dense Prediction for 360° Depth Estimation

Abstract

The equirectangular projection (ERP) is widely used for panoramic depth estimation, but its spatially varying distortion makes geometry-consistent feature modeling challenging. We revisit panoramic depth estimation by decoupling contextual modeling in native spherical space from dense ERP prediction. To this end, we propose a Fibonacci Spherical Graph (FSG) as an intermediate reasoning space to lift ERP features onto quasi-uniform Fibonacci nodes on the sphere and capture local and long-range dependencies through complementary spherical neighborhoods. The resulting spherical discretization distributes graph nodes approximately uniformly over the spherical surface, reducing the over-representation of highly stretched regions during relational modeling. Operating on a compact set of Fibonacci nodes also avoids the computational burden of constructing and processing a graph at full ERP resolution. To bridge spherical reasoning and dense prediction, we propose a Spherical Context Conditioning (SCC) module that adaptively modulates dense ERP features with the enhanced spherical representation, allowing spherical context to guide pixel-aligned depth prediction. Extensive experiments on three benchmarks demonstrate that the proposed method consistently achieves superior depth accuracy over existing approaches.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
360 vision

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nuiEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Proposes a panoramic depth estimation framework that separates spherical context reasoning from dense ERP prediction. Uses a Fibonacci Spherical Graph (FSG) for geometry-aware feature aggregation and Spherical Context Conditioning (SCC) to integrate spherical features into ERP representations. Achieves improved depth accuracy on three benchmarks.

Critical strengths
  1. Separating spherical reasoning from dense ERP prediction is a reasonable approach to handling projection distortion.

  2. Achieves the best results across three panoramic depth benchmarks. Ablations support the contributions of FSG and SCC.

Critical weaknesses
  1. Limited methodological novelty. Spherical graph processing and ERP feature fusion have been explored in previous works. The main distinction from methods such as Elite360D needs clearer justification
  2. Unclear contribution of graph relations. Adding dilated connections to local relations alone degrades performance (Table 3). The benefit of the proposed multi-relation design needs further explanation.
Questions for authors
  1. What is the key methodological difference from Elite360D and other spherical representation-based methods?
  2. Why does adding dilated connections degrade performance without antipodal connections (Table 3)?
  3. How sensitive is performance to the number and distribution of Fibonacci nodes?
  4. Can the authors clarify the accuracy–latency trade-off relative to UniFuse?
Reviewer_BStYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel bi-projection fusion-based 360 depth estimation framework, which leverages the Fibonacci Spherical Graph to acheive the more uniform shperical discretization and tackle the distortion issue of direct dense ERP format predictions. The addtional SCC module supports the information interaction inside the ERP and improve the spherical reasoning. The mathematic formulations in the main paper and supplementary material clearly provide the implementation details. And the experiments show the superior performance compared with the current bi-projection fusion-based methods.

Critical strengths
  1. The design of decoupling the dense prediction and spherical reasoning is well-motivated. Compared the current bi-projection fusion-based methods, this work explores the potential of intra-projection information fusion and subtly uses the Fibonacci Spherical Graph to achieve the uniform discretization.
  2. The experiments and ablation studies demonstrate the effectiveness of each module.
  3. The paper is well-structured and easy to follow.
Critical weaknesses

Honestly, I have several heavy concerns about the technology.

  1. More claim about the cubemap branch. Basically, the FSG and SCC modules already provide the feature fusion between distorted original ERP features X and the uniformly sampled spherical context G. The start point of UniFuse is to combined the less-distored cubmap features with globally perceived but distorted ERP features. Why the authors still employ the cubemap branch?

  2. More claim about the FSG.While the use of Fibonacci sampling is mathematically beautiful, its necessity remains questionable. Table 2 only provides the comparison between the Fibonacci support and ERP-grid, however, there are several approximately uniform spherical discretizations. More experiments to clarify if the improvements are brought by Fibonacci sampling is beneficial.

  3. More comparison with 360 depth estimation methods, which are based on foundation models. I understand why the authors mainly compare the proposed method with the bi-projection fusion-based methods. However, depth estimation, especially 360 depth estimation, is a vision task most aiming for the downstream application, so the performance is much important. Currently, there are so many depth-based foundation models and the related works, such as PanDA, ST²360D, Depth Any Camera, Depth Any Panoramas,DA2. The performance comparison could verify the proposed method’s practical competitiveness.

Minor

  1. More experiments based on the self-captured data or in-the-wild data could be better to verify the robustness.

[1] Cao, Zidong, et al. "PanDA: Towards panoramic depth anything with unlabeled panoramas and Möbius spatial augmentation." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. [2]Cao, Zidong, et al. "ST $^ 2 $360 D: Spatial-to-Temporal Consistency for Training-free 360 Monocular Depth Estimation." Advances in Neural Information Processing Systems 38 (2026): . [3]Guo, Yuliang, et al. "Depth any camera: Zero-shot metric depth estimation from any camera." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. [4]Lin, Xin, et al. "Depth any panoramas: A foundation model for panoramic depth estimation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026. [5] Li, Haodong, et al. "DA $^{2} $: Depth Anything in Any Direction." arXiv preprint arXiv:2509.26618 (2025).

Questions for authors

Please see the weakness, and I strongly suggest that the authors could provide the more experiments to show the necessity of bi-projection fusion and compare with more 360 depth estimation methods based on the foundation model.

Reviewer_HmjwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper decouples contextual modeling in native spherical space from dense ERP prediction. Specifically, they incorporate a FSG and a SCC module into the UniFuse dual-projection pipeline across four feature scales. FSG serves as a compact intermediate reasoning space that lifts ERP features onto quasi-uniform Fibonacci nodes on the sphere and captures local, intermediate-range, and long-range dependencies through complementary local, dilated, and antipodal spherical neighborhoods. To bridge spherical reasoning and dense prediction, the SCC module keeps the ERP feature as the carrier of dense information and adaptively modulates it with the back-projected spherical context through gated affine modulation and context injection, allowing spherical context to guide pixel-aligned depth prediction. Extensive experiments and controlled component ablations on three indoor benchmarks demonstrate that the proposed method consistently achieves superior depth accuracy over existing approaches while maintaining lower computational overhead than recent competing methods.

Critical strengths
  • Well-motivated decoupling via compact FSG: this paper decouples contextual reasoning from dense prediction effectively alleviates latitude-dependent distortion and sampling imbalance. Operating on a compact set of quasi-uniform Fibonacci nodes reduces the over-representation of highly stretched polar regions and assigns more effective representation capacity to equatorial regions, while avoiding the computational burden of constructing and processing a graph at full ERP resolution. Consequently, the model requires fewer FLOPs (104G), less training memory (3.37 GB), and lower inference latency (38ms) than recent methods achieving comparable accuracy such as PGFuse and HUSH.

  • Complementary multi-relation spherical modeling and SCC: Constructing local, dilated, and antipodal relations over the same spherical nodes allows nearby viewing directions and direct communication with the opposite part of the panorama to provide complementary structural context. Furthermore, instead of making pixel-level detail depend solely on interpolated graph representations, SCC adaptively conditions dense ERP features with graph-derived spherical context via gated modulation, enabling selective information transfer that improves large-scale structural consistency while preserving local geometric detail.

  • Strong empirical validation: The method achieves the lowest depth errors and highest threshold accuracies across all three evaluated benchmarks.

Critical weaknesses
  • Unexamined functional overlap with the Cubemap branch in the dual-projection baseline: The primary motivation for introducing the FSG is to overcome the latitude-dependent distortion and inconsistent neighborhood structures of ERP. However, the proposed method is built entirely upon UniFuse, retaining its dual-projection ERP and Cubemap encoders, cross-projection fusion modules, and ERP decoder. Because Cubemap representations are specifically designed to complement ERP features and mitigate severe polar projection distortion, retaining the Cubemap branch makes it difficult to disentangle the distortion-handling contribution of FSG from that of the auxiliary Cubemap stream. The paper omits a critical ablation evaluating FSG and SCC on a standalone, single-branch ERP network without Cubemap fusion.

  • Incremental empirical gain from Fibonacci node support over uniform ERP-grid graph reasoning: While quasi-uniform Fibonacci discretization is emphasized as a core contribution to avoid high-latitude over-representation, the controlled comparison in Table 2 indicates that the majority of the performance improvement comes from the updated training objective and the general addition of graph aggregation and gated conditioning rather than the Fibonacci node distribution itself.

  • Limited architectural validation on a single legacy backbone: Although decoupling spherical reasoning from dense ERP prediction is framed as a general formulation for panoramic depth estimation, all experiments only instantiate FSG and SCC within a single ResNet-34-based UniFuse pipeline from 2021. The paper lacks empirical verification demonstrating that the proposed modules generalize effectively to pure ERP-based baselines or stronger modern vision backbones.

Questions for authors
  1. Since UniFuse already relies on a Cubemap encoder to compensate for ERP polar distortion, how do FSG and SCC perform when integrated into a pure single-branch ERP encoder-decoder pipeline? Reporting this ablation on Matterport3D would materially clarify whether FSG can independently resolve ERP distortion and potentially eliminate the computational overhead of maintaining a dual-projection encoder.

  2. In Table 2, ERP Grid + Graph + SCC already achieves most of the performance gain over the baseline, leaving a relatively small margin when switching to Fibonacci support. Could the authors clarify whether the retained Cubemap branch in UniFuse is already compensating for the high-latitude redundancy of the ERP Grid variant, and whether a larger gap between ERP Grid and Fibonacci support appears when evaluated without the Cubemap branch or on the other two benchmarks?

  3. Have the authors evaluated plugging FSG and SCC into a pure ERP architecture or a more recent backbone beyond ResNet-34 UniFuse?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.