acceptodds
Under review as a conference paper at ICLR 2027

Disentangled Panorama Super-Resolution with Latitude-Wise Modeling

Abstract

Panorama super-resolution suffers from entangled content and degradation features, as well as latitude-dependent distortion caused by equirectangular projection. Existing end-to-end methods jointly learn mixed features, limiting their generalization and amplifying artifacts in high-latitude regions. We propose a decoupled dual-branch framework that explicitly separates sample-specific content features from universal degradation patterns. Two dedicated encoders are trained with a cross-sample feature fusion strategy, forcing the degradation branch to capture only content-agnostic information while the content branch focuses on semantic and geometric structures. To handle non-uniform stretching, we design a latitude-wise block with adaptive convolution whose dilation rate is dynamically modulated by the spherical polar angle. Edge-guided spatial attention and a composite loss with feature consistency further enhance detail recovery and training stability. Our method achieves state-of-the-art performance on standard benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Panoramic Image Super-Resolution, Panoramic Image, Computer Vision

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_u7qiOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies super-resolution for panoramic images represented in the equirectangular projection (ERP) format. It proposes:

  1. A dual-branch architecture consisting of a content encoder and a degradation encoder.
  2. Cross-sample swapping of degradation features during training to encourage content–degradation disentanglement.
  3. A latitude-wise convolution module that interpolates between convolutions with dilation rates ss and s+1s+1, using ∣cos⁡θ∣|\cos\theta| as the latitude-dependent weight.
  4. Channel attention, edge-guided spatial attention, and a training objective combining L1, SSIM, and feature-consistency losses.

On ODI-SR and SUN360, the proposed method improves WS-PSNR over EAM by approximately 0.09–0.25 dB while using fewer parameters.

Critical strengths
  1. The paper studies a practically relevant issue in ERP panoramic SR: non-uniform distortion across latitudes. I think this is a very meaningful perspective for research.
  2. The proposed network contains only 1.3M parameters and achieves a favorable trade-off between model size and reconstruction accuracy.
Critical weaknesses

1. Cross-sample swapping does not establish disentanglement and admits trivial solutions

The method concatenates FC,1F_{C,1} with FD,2F_{D,2}—a degradation representation extracted from a different image—and supervises the resulting reconstruction using the HR target corresponding to the first image. It additionally minimizes the cosine dissimilarity between degradation features extracted from different samples.

However, this objective admits at least two trivial solutions:

  • The degradation encoder may collapse to an approximately constant representation for all samples sharing the same bicubic degradation.
  • The decoder may simply ignore FDF_D and rely entirely on FCF_C for reconstruction.

The evidence presented in Section 4.5 and Appendix A.5 does not rule out either possibility and, in my view, is particularly consistent with the second. Because the training data are generated primarily using a fixed bicubic degradation, the content encoder can encode not only image content but also all information needed to invert the degradation. The proposed framework contains no explicit mechanism—such as an information bottleneck, an adversarial content classifier, a mutual-information constraint, or a controlled intervention objective—to prevent such information leakage from the degradation branch into the content branch.

The appendix reports that cross-sample swapping of degradation features leaves PSNR unchanged and interprets this result as direct evidence of successful disentanglement. This interpretation is not justified. Invariance to swapping is equally consistent with FDF_D being constant, uninformative, or ignored by the decoder. Therefore, the reported swapping experiment does not demonstrate that FDF_D captures meaningful, content-independent degradation information. On the contrary, without additional intervention-based ablations—such as zeroing, randomizing, or replacing FDF_D with its dataset mean—the result raises the concern that the degradation branch has little or no causal influence on reconstruction.

2. The definition of “sample-agnostic degradation features” is internally inconsistent

The paper claims that degradation features are sample-agnostic and can be exchanged across images. At the same time, it states that these features contain spatial responses associated with local gradients and edges.

For bicubic downsampling, local degradation residuals depend strongly on the specific image content, frequency distribution, and edge locations. If FD(A)F_D(A) preserves the spatial degradation responses of image AA, it is unclear why it should be spatially compatible with the content representation of image BB. Conversely, if swapping FD(A)F_D(A) and FD(B)F_D(B) makes no difference, this suggests that FDF_D contains only a global degradation code or is ignored altogether.

Questions for authors
  1. How do the authors rule out the possibility that the decoder ignores the degradation branch?
  2. If degradation features contain spatial patterns tied to edge locations, why should features from image AA help reconstruct image BB?
Reviewer_ssbZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a panorama super-resolution method with two main components.

First, it uses separate content and degradation encoders. During training, degradation features are swapped across different samples before being fused with content features, with the intention of making the degradation representation content-independent.

Second, it introduces a latitude-wise convolution whose dilation varies with ERP latitude, aiming to enlarge the receptive field toward the poles where projection distortion is stronger. Channel attention, edge-guided spatial attention, and a composite reconstruction loss are also included.

The model is trained on an augmented ODI-SR dataset and evaluated on ODI-SR and SUN360 for ×8 and ×16 super-resolution.

Critical strengths
  1. The paper addresses relevant characteristics of omnidirectional SR, including latitude-dependent ERP distortion and degradation modeling.

  2. The proposed components do provide measurable improvements. In particular, the latitude-wise convolution improves WS-PSNR by about 0.24 dB, with a somewhat larger gain in high-latitude regions.

  3. The model has a small parameter count and consistently improves over EAM, the most recent baseline using the same enlarged training protocol.

Critical weaknesses

1. The central content/degradation disentanglement claim is not convincingly established.

The proposed cross-sample feature swapping encourages the degradation branch to be invariant to sample identity, but sample invariance does not imply that the learned feature specifically represents degradation.

More importantly, the current experiments do not rule out trivial explanations. If the decoder relies primarily on the content feature and is only weakly sensitive to the degradation feature, then swapping degradation features across samples would naturally leave the reconstruction almost unchanged. Similarly, a nearly constant degradation representation could also satisfy the cross-sample consistency constraint.

The DFeat ablation in Table 3 does not resolve this issue. According to the paper, this ablation removes the cross-sample fusion constraint while keeping the dual-branch architecture and parameter count unchanged. Therefore, it demonstrates that the cross-sample training strategy is useful, but does not demonstrate that the decoder actually needs or meaningfully uses the degradation branch.

A direct branch-sensitivity experiment is missing. For a trained model, the authors should test what happens when the degradation feature is zeroed, replaced by its dataset mean, randomized, or when the degradation encoder is removed entirely.

The additional evidence in Appendix A.5 is also difficult to interpret. The bicubic and salt-and-pepper models are trained independently, yet their latent features are directly compared using cosine similarity. Since independently trained latent spaces are not aligned, a cross-model cosine similarity of 0.66 for content and 0.10 for degradation cannot by itself establish disentanglement.

Overall, the experiments show that the proposed architecture/training strategy improves SR performance, but they do not establish the stronger claim that the two branches have learned identifiable content and degradation factors.

2. The degradation claim is evaluated almost entirely under one synthetic bicubic degradation.

The main ODI-SR/SUN360 setting generates LR inputs from HR panoramas using synthetic bicubic downsampling. Thus, although the underlying panoramas are real images, the degradation used for the main training and evaluation is not a real camera degradation.

This is particularly problematic because the paper motivates the degradation branch as learning "universal degradation patterns" and argues that disentanglement improves generalization. Under the main training setting, however, almost every sample follows essentially the same known bicubic degradation rule. There is insufficient degradation variation to establish such a claim.

The salt-and-pepper experiment does not provide an unseen-degradation test of the same trained model; a separate model is trained for that degradation.

This limitation is important in the current literature. For example, OSRT already introduced fisheye downsampling to better approximate the real omnidirectional imaging pipeline, rather than directly applying bicubic downsampling in ERP space. More recently, the following closely related work explicitly studies real-world omnidirectional degradation disentanglement and should be discussed:

  • Hongyu An, Xinfeng Zhang, Xu Fan, Shijie Zhao, Li Zhang, and Ruiqin Xiong. "D²R²OSR: Degradation-Disentangled Representation for Real-World Omnidirectional Image Super-Resolution." European Conference on Computer Vision (ECCV), 2026.

D²R²OSR explicitly models real-world degradations such as blur, noise, resizing, and compression together with projection-related degradation. This makes the current paper's broad degradation-disentanglement and generalization claims difficult to justify using only the conventional bicubic benchmark.

3. The latitude-wise convolution is useful, but its novelty and geometric justification are limited.

Latitude-aware processing is already well established in omnidirectional SR. LAU-Net adapts processing according to latitude, while OSRT introduces latitude-conditioned distortion-aware attention and convolution. Therefore, the broad idea of adapting the receptive field according to ERP latitude is not new.

The specific contribution here is narrower: blending two standard convolutions with adjacent integer dilation rates using a smooth latitude-dependent weight. This is a lightweight implementation and does provide a measurable gain, but it is an incremental design rather than a new panorama-SR principle.

There is also a problem with the stated geometric derivation. The paper defines θ\theta as the zenith angle, with the poles at 0∘/180∘0^\circ/180^\circ and the equator at 90∘90^\circ, and then states that horizontal ERP stretching is proportional to ∣cos⁡θ∣|\cos\theta|. For a sphere under this parameterization, however, the physical horizontal distance associated with a longitude increment is proportional to sin⁡θ\sin\theta, and the relative ERP stretching toward the poles is related to 1/sin⁡θ1/\sin\theta, not ∣cos⁡θ∣|\cos\theta|.

Therefore, ∣cos⁡θ∣|\cos\theta| is better interpreted as a bounded heuristic that monotonically increases toward the poles, rather than the actual spherical distortion ratio.

It is also unclear whether the dilation is applied horizontally only or isotropically in both spatial dimensions. Since the latitude-dependent ERP distortion considered here is mainly horizontal, isotropic dilation would require additional justification.

4. The overall quantitative gain is modest relative to the strength of the claims.

Compared with the matched EAM baseline, the improvements are approximately 0.09/0.14 dB for ×8 and 0.21/0.25 dB for ×16 on ODI-SR/SUN360. These are consistent improvements, but relatively modest.

Although the proposed model has fewer parameters than EAM (1.3M vs. 2.0M), it requires substantially more computation (106G vs. 38G FLOPs) and is slightly slower at inference (0.030 s vs. 0.022 s). Therefore, the claimed lightweight property mainly applies to parameter count rather than computational efficiency.

In addition, most older baseline results in Table 1 are taken from the original papers, while the proposed method is trained using the enlarged 12,000-sample spherical augmentation protocol introduced by EAM. The authors appropriately identify EAM as the primary matched baseline, but the full table should not be interpreted as a strictly controlled comparison across all methods.

Given that the strongest conceptual contribution, degradation disentanglement, is not convincingly validated, I do not think these empirical gains alone are sufficient to support the paper's current claims.

Questions for authors
  1. Does the decoder actually depend on the degradation feature?
    For a trained full model, what happens if FDF_D is zeroed, replaced with its dataset mean, randomized, or removed entirely? The current DFeat ablation only removes the cross-sample constraint while retaining the degradation branch, so it does not rule out the possibility that the decoder mainly relies on FCF_C.

  2. Can the authors provide a controlled disentanglement experiment using a single model trained on multiple degradation types or strengths?
    For example, content/degradation features could be cross-swapped across different degradations, and linear probes could measure how much content and degradation information is present in each branch.

  3. How does the method generalize to real or unseen degradations?
    The main model is trained and evaluated under synthetic bicubic downsampling. Please evaluate the same trained model under realistic fisheye/ERP degradations, blur, noise, compression, or mixed degradations rather than training a separate model for each degradation.

  4. Can the authors clarify the geometry behind the ∣cos⁡θ∣|\cos\theta| dilation rule?
    With the zenith-angle definition used in the paper, the horizontal spherical metric depends on sin⁡θ\sin\theta. Why is ∣cos⁡θ∣|\cos\theta| described as the ERP stretching ratio? A comparison with a Jacobian- or distortion-map-based horizontal scaling would be useful.

  5. Is latitude-dependent dilation applied horizontally only or in both dimensions?
    If it is isotropic, what is the justification for increasing the vertical receptive field as well?

  6. Can the authors discuss recent real-world degradation-disentangled ODISR methods, especially D²R²OSR (ECCV 2026)?
    The current novelty statement should distinguish the proposed synthetic-bicubic feature separation from recent methods that explicitly model projection and real-world degradations.

Reviewer_3nYyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a dual-branch framework for panorama image super-resolution. It uses separate content and degradation encoders, with cross-sample feature fusion to encourage the degradation branch to learn content-independent features. The method also introduces latitude-wise convolutions and attention modules to address ERP distortion and recover details. Experiments on ODI-SR and SUN360, including component and latitude-region ablations, show improvements over the reported baselines.

Critical strengths
  1. The problem is clearly defined, and the overall approach is intuitive. The paper clearly explains the challenges of super-resolution for ERP panoramas, and the proposed design provides a straightforward way to address them.

  2. The experiments include useful component analyses. The ablations and latitude-region results help assess whether the proposed modules contribute to performance.

Critical weaknesses
  1. The methodological novelty is unclear. The proposed method appears to draw on ideas from BPOSR (AAAI 2024) and EAM (CVPR 2026), while its problem formulation and central ideas appear to overlap substantially with D²R²OSR (ECCV 2026). The paper should clarify its novel contributions and distinguish them from these related methods.

  2. The geometric justification for the latitude modulation could be strengthened. The paper motivates latitude-dependent dilation using ERP distortion, but does not derive how the chosen bounded latitude weight follows from the ERP scale factor or why it is an appropriate proxy. Please clarify this choice and explain how latitude-weighted convolutions differ from spherical convolution in their geometric treatment and feature aggregation.

  3. The evaluation omits ×4 results and provides limited evidence of perceptual gains. The paper reports results only for ×8 and ×16 upscaling, with no ×4 evaluation, a commonly reported ODISR setting in methods such as BPOSR (AAAI 2024), OSRT (CVPR 2023), and LAU-Net (CVPR 2021). The visual comparisons also show only limited perceptual improvement over EAM (CVPR 2026).

  4. The end-to-end evaluation focuses on bicubic degradation. The appendix analyzes degradation features under salt-and-pepper noise, but does not report full restoration results under that or other degradation types. It therefore remains unclear how well the method generalizes beyond bicubic downsampling to more realistic or varied degradations. Real-world ODI-SR has been explored in Real-OSR (ECCV 2024) and D²R²OSR (ECCV 2026).

Questions for authors
  1. Were the baseline methods evaluated under comparable training protocols? The proposed method is trained using the data augmentation protocol adopted by EAM, while Table 1 reports results for other methods from their original papers. Please clarify the training data and evaluation protocol used for each baseline, and whether the baselines were retrained under a common protocol.
  2. How is the claimed improvement in perceptual quality supported? The quantitative evaluation mainly reports WS-PSNR and WS-SSIM, while the visual comparisons are limited. Could the authors provide additional perceptual metrics, such as LPIPS or DISTS, to better substantiate the claimed visual improvements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.