acceptodds
Under review as a conference paper at ICLR 2027

Blind Omnidirectional Video Super-Resolution via Geometry-Conditioned Probabilistic Degradation Modeling

Abstract

Blind Video Super-Resolution (BVSR) promotes generalization across diverse imaging conditions by explicitly accounting for unknown degradations to guide reconstruction. However, its extension to OmniDirectional Video (ODV) remains underexplored. The unique fisheye imaging and EquiRectangular Projection (ERP) process of ODVs couples unknown capture degradations with projection geometry, making local degradation characteristics jointly dependent on both factors and complicating degradation inference. Therefore, we propose OmniVaria, the first blind ODVSR framework with a particular focus on modeling the interaction between unknown degradations and spherical geometry. In particular, we introduce Probabilistic Spherical Degradation Conditioning (PSDC), which leverages explicit spherical geometry as an additional condition in probabilistic degradation modeling, i.e., representing unknown degradations through geometrically conditioned multivariate Gaussian distributions in a latent space, with posterior statistics inferred from Low-Resolution (LR) ODV observations guiding reconstruction. For PSDC optimization, we establish an information-theoretically grounded regularization term, named Geometry-Conditioned Variational Information Bottleneck (GCVIB), and derive its tractable variational form for practical optimization. Drawing on information bottleneck theory, GCVIB encourages PSDC to learn a minimally sufficient representation of LR observations for High-Resolution (HR) reconstruction, thereby promoting generalization to unseen scenes and degradation levels. Furthermore, recognizing that ERP projection transforms coherent scene motion into spatially varying displacements and deformations, complicating the exploitation of temporal context, we bypass explicit motion estimation and develop Implicit Omnidirectional Temporal Aggregation (IOTA), which jointly processes current-frame features and propagated hidden states to exploit inter-frame correlations for reconstruction. Extensive experiments on various ODV benchmarks and degradation settings demonstrate the effectiveness of our design. The code will be publicly released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
omnidirectional video super-resolution, blind video super-resolution

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_RXUkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes OmniVaria, a blind omnidirectional video super-resolution framework that models the interaction between unknown degradations and spherical geometry. It introduces probabilistic degradation conditioning with an information bottleneck regularizer, along with implicit temporal aggregation without explicit motion estimation. Experiments on three benchmarks show promising performance, particularly under unseen blur levels.

Critical strengths

1.The paper addresses a meaningful and relatively underexplored problem by extending blind video super-resolution to omnidirectional videos.

2.The paper makes an accurate observation that fisheye imaging and ERP projection jointly lead to spatially varying degradations and complicate temporal correspondence in omnidirectional videos.

Critical weaknesses

1.Several core ideas, including implicit degradation representation learning, separating degradation information from image content, and information bottleneck regularization, have already been explored in blind image super-resolution. While adapting these ideas to omnidirectional videos is meaningful, the methodological contributions beyond geometry-aware extensions of existing approaches could be better clarified.

2.The evaluation is limited to a single synthetic fisheye degradation pipeline with isotropic Gaussian blur and downsampling. Although the model generalizes to unseen blur levels, its effectiveness under other degradation types or realistic capture conditions remains unclear.

Questions for authors

1.Since both degradation representation learning and information bottleneck regularization are well-studied in blind SR, please clarify the specific methodological novelty of this framework over existing approaches.

2.How does the proposed GCVIB formulation fundamentally differ from standard IB methods? It would be helpful to provide more concrete evidence showing how it suppresses image content in the learned degradation representations.

3.The generalization experiments only vary blur levels within a fixed degradation process. Can the model handle entirely different degradation types beyond isotropic Gaussian blur?

4.All current evaluations rely on synthetic data. Have the authors tested the model on real-world omnidirectional videos with more complex, natural degradations?

2 comments

First, we sincerely appreciate your thorough and constructive review, which helped improve the manuscript and inspired our future work. Regarding your concerns, we hope the following responses address them, and we would appreciate any further suggestions.

W.1 & Q.1 Methodological Contribution

Ans.: Thanks for your constructive comments. We would like to emphasize that OmniVaria follows a unified geometry-driven design to address the coupled spatial and temporal challenges of blind ODVSR.

Specifically, fisheye imaging and ERP reprojection transform unknown capture degradations into spatially varying distortions. Rather than merely compensating for geometric distortion, PSDC uses explicit spherical geometry to guide unknown degradation inference, producing latitude-structured Gaussian conditions that enable spatially varying, posterior-statistics-guided feature modulation. GCVIB further regularizes the information capacity of these conditions while preserving reconstruction-relevant information, potentially alleviating over-specialization to training degradation statistics. Meanwhile, projection-dependent motion deformation complicates temporal correspondence, motivating IOTA to aggregate recurrent features without explicit flow estimation or warping.

Thus, geometry explicitly guides degradation conditioning and motivates temporal aggregation, forming a coherent reconstruction framework. Our geometry-conditioning, local-condition, and temporal-alignment ablations further substantiate these design choices.

W.2 & Q.3 Generalization Beyond Gaussian Blur

Ans.: Thanks for your thorough comments which made us realize the importance of generalization. To this end, we conduct zero-shot evaluation on three unseen degradation families using ODV360.

DegradationLevelOmniVariaFMA-NetSTDANOSRTw/o Prob.w/o GCVIB
Anisotropic Gaussian Blur(ηmaj, ηmin)WS-PSNR
(1.8, 0.6)26.1924.9925.4225.5024.8925.74
(2.4, 0.8)26.0924.8725.3825.2824.8125.63
(3.5, 0.8)25.6024.5124.8224.7324.3725.16
Gaussian Noiseη
525.4423.9823.1424.8724.0725.04
1024.0722.7122.0723.5822.6323.77
2021.5921.1720.1621.0720.2721.24
H.264 CompressionCRF
1825.6924.7724.6925.1624.3825.27
2325.2424.5024.5124.8224.4824.91
2824.4923.9724.0024.2523.9924.28

As shown above, OmniVaria consistently outperforms the baselines, showing promising generalization capability toward unknown degradations.

Q.2 GCVIB Formulation and Content Suppression

Ans.: Thank you for this insightful comment, which helped us realize that our phrasing regarding the role of GCIVB was inappropriate. GCVIB builds upon standard IB regularization but differs due to its geometry-conditioned, latitude-structured probabilistic representations that directly guide spatially varying HR reconstruction.

Specifically, the KL regularizer provides a variational upper bound on I(Z;X∣C)I(Z;X\mid C), constraining the information retained by the inferred conditions from LR observations given the known ERP geometry. Meanwhile, the reconstruction-relevance term I(Z;Y∣F,C)I(Z;Y\mid F,C) motivates preserving information complementary to the existing reconstruction features and geometric prior. This regularization of information compression and reconstruction relevance is particularly important for blind ODVSR. Without such regularization, GCDI may overfit to the degradation statistics during training, potentially reducing the reliability of feature adaptation under unseen degradations, as demonstrated in W.2 & Q.3.

Therefore, we reposition GCVIB's role as a regularization of probabilistic degradation conditioning and promote generalization capability through the joint optimization of the compression and reconstruction objectives. We have revised the manuscript to align our description with the mathematical and experimental support.

Q.4 Evaluation on Real-world Omnidirectional Videos

Ans.: Thanks for your thorough comments. Evaluating blind ODVSR under natural capture degradations remains challenging but meaningful. While existing ODVSR benchmarks concentrate on constructing LR frames through synthetic degradations, i.e., fisheye downsampling and HEVC compression. In the current stage, we derive a zero-shot evaluation on synthetic degradations as an alternative assessment. Your valuable suggestions have motivated us to further develop a real-world ODVSR benchmark containing more complex, natural degradations in our future work.

Currently, we are trying our best to incorporate these valuable results and method repositioning into the revised manuscript, and we are pleased to address any further questions.

Thank you for the authors' detailed response and additional experiments; however, my main concerns regarding methodological novelty and real-world applicability remain insufficiently addressed, so I will maintain my original rating.

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

This paper proposes OmniVaria for blind omnidirectional video super-resolution. It uses spherical geometry to infer local probabilistic degradation conditions and guide feature restoration. The model also includes an information bottleneck regularizer and a temporal aggregation module that combines features across frames without explicit optical flow estimation. Experiments on multiple datasets show improvements under varying Gaussian blur strengths, including strengths not seen during training.

Critical strengths

S1. The paper addresses blind omnidirectional video SR, which is a challenging and relatively underexplored problem.

S2. The proposed framework is well motivated. It considers both spatially varying degradations and temporal information from neighboring frames.

S3. The method achieves good results across three datasets, especially on unseen blur strengths. The ablation studies also support the effectiveness of the main components.

Critical weaknesses

W1. A central contribution of OmniVaria is PSDC, which uses spherical geometry to infer probabilistic degradation conditions and guide restoration. However, D²R²OSR [1] has already explored geometry-aware degradation modeling for omnidirectional image restoration, considering diverse degradations such as blur, noise, resizing, and compression. While PSDC introduces a probabilistic formulation, the benefits of probabilistic modeling need further analysis. Moreover, the evaluation mainly considers isotropic Gaussian blur with varying strengths, which may limit the method's practical usefulness.

W2. The role of GCVIB. Although based on an information bottleneck objective, GCVIB uses a standard KL regularization term. Removing GCVIB causes only a small performance drop (0.06 dB on ODV360), making it unclear whether GCVIB can reduce scene-specific information while keeping useful degradation information.

[1] An H, Zhang X, Fan X, et al. D²R²OSR: Degradation-Disentangled Representation for Real-World Omnidirectional Image Super-Resolution[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026: 1-19.

Questions for authors
  1. What are the specific benefits of probabilistic degradation modeling in PSDC compared with deterministic geometry-aware degradation modeling?

  2. Table 4 shows that OmniVaria has relatively high inference latency. The authors should provide further analysis of the reasons behind this overhead.

3 comments

First, we sincerely appreciate your thorough and constructive review. Regarding your concerns, we hope the following responses address them, and we would appreciate any further suggestions.

W.1 & Q.1 Novelty of OmniVaria and Benefits of Probabilistic Degradation Modeling

Ans.: Thank you for your valuable suggestions. We would like to emphasize that both D²R²OSR and our OmniVaria consider the interplay between panoramic imaging geometry and degradation. However, they differ fundamentally in objectives and mechanism.

Specifically, D²R²OSR targets real-world ODISR by explicitly disentangling projection distortions and degradations via complementary ERP/PPR representations and degradation-specific modules. In contrast, OmniVaria addresses blind ODVSR, where fisheye imaging and ERP reprojection jointly induce spatially varying degradation characteristics and complicate temporal correspondence.

To address this, PSDC jointly infers latitude-structured Gaussian posteriors from LR observations and ERP geometry prior. Unlike deterministic point-estimate conditioning, the probabilistic representation provides both posterior means and variances: the means serve as continuous reconstruction conditions, while the variances provide additional signals for adaptive feature gating. This design enables degradation-adaptive reconstruction without requiring explicit degradation type identification. GCVIB further regularizes the information capacity of these conditions, while IOTA exploits temporal information without explicit optical-flow estimation or warping.

To verify the benefits of probabilistic modeling and generalization beyond isotropic Gaussian blur, we conduct zero-shot evaluations on ODV360 under three typical, unseen degradation families.

DegradationLevelOmniVariaFMA-NetSTDANOSRTw/o Prob.w/o GCVIB
Anisotropic Gaussian Blur(ηmaj, ηmin)WS-PSNR
(1.8, 0.6)26.1924.9925.4225.5024.8925.74
(2.4, 0.8)26.0924.8725.3825.2824.8125.63
(3.5, 0.8)25.6024.5124.8224.7324.3725.16
Gaussian Noiseη
525.4423.9823.1424.8724.0725.04
1024.0722.7122.0723.5822.6323.77
2021.5921.1720.1621.0720.2721.24
H.264 CompressionCRF
1825.6924.7724.6925.1624.3825.27
2325.2424.5024.5124.8224.4824.91
2824.4923.9724.0024.2523.9924.28

As shown, OmniVaria shows pronounced generalization capability. We attribute this to the collaboration of PSDC and GCVIB. Replacing them with deterministic conditions results in a dramatic performance decline, while GCVIB provides a further gain beyond PSDC.

W.2 Clarifying and Repositioning the Role of GCVIB

Ans.: Thanks for your insightful comments, which helped us realize our original description of suppressing scene-specific information was inappropriate given the available evidence, and we will revise it accordingly.

GCVIB builds upon the standard variational information bottleneck but differs in applying conditional information regularization to geometry-conditioned, latitude-structured probabilistic representations. Specifically, the KL-based regularizer provides a variational upper bound on I(Z;X∣C)I(Z;X\mid C), constraining the information retained from LR observations given ERP geometry. Meanwhile, the reconstruction-relevance term I(Z;Y∣F,C)I(Z;Y\mid F,C) encourages the preservation of complementary information useful for HR recovery beyond the existing reconstruction features and geometric prior. This balance between information compression and reconstruction relevance may alleviate over-specialization to training degradation statistics, thereby improving robustness under distribution shifts. Importantly, our additional zero-shot results in W.1 & Q.1 demonstrate this generalization benefit.

We therefore reposition GCVIB as an information-capacity regularizer for geometry-conditioned probabilistic degradation representations and have revised the corresponding content in the manuscript.

Q.2 Detailed Analysis of Inference Latency

Ans.: Thanks for your careful review. Herein, we deliver an additional module-wise latency analysis of OmniVaria.

Componentms/frame
Transformer(excl. DCFM)1389.35
DCFM83.04
IOTA7.57
GCDI0.43

The computational bottleneck lies in the recurrent Transformer backbone, whereas our proposed GCDI, DCFM, and IOTA introduce minimal overhead.

Currently, we are trying our best to incorporate these valuable results and method repositioning into the revised manuscript, and we are pleased to address any further questions.

Thanks for the response and additional experiments. The new results are helpful and address some of my concerns. However, I'm still not fully convinced by the novelty of the method. In particular, it's unclear whether the gains come from probabilistic degradation modeling itself or simply from the additional feature conditioning and gating. Overall, I will keep my rating unchanged.

Thank you for reading our responses and for raising your concerns. To more directly disentangle the contributions of probabilistic modeling and feature gating, we have just completed the full training of an ablated model with deterministic conditioning and gating. It adopts an MLP to project the deterministic condition to a gating signal. Together with OmniVaria and w/o Prob. variant, a more comprehensive comparison under various degradation families is reported below.

DegradationLevelOmniVariaw/o Prob. w/o Gatew/o Prob. w/ Gate
Isotropic Gaussian Blurη\etaWS-PSNR
0.026.1625.7925.87
1.226.2525.9426.00
1.826.2525.9626.02
2.426.2025.9025.91
3.525.5425.0624.93
Anisotropic Gaussian Blurηmaj\eta_{maj}, ηmin\eta_{min}
1.8, 0.626.1924.8924.95
2.4, 0.826.0924.8124.86
3.5, 0.825.6024.3724.37
Gaussian Noiseη
525.4424.0724.16
1024.0722.6322.75
2021.5920.2720.31
H.264 CompressionCRF
1825.6924.3824.73
2325.2424.4824.51
2824.4923.9924.04

As shown, introducing gating to deterministic conditioning yields only marginal improvements. More importantly, our previous w/o GCVIB retaining probabilistic conditioning and variance gating further improves the performance.

Empirically, in the gated deterministic variant, conditions take responsibility for feature modulation and gating, which is naturally disentangled by different distribution statistics in probabilistic conditioning. We attribute the further performance gain to the collaboration of PSDC and GCVIB, and thank you again for the detailed discussion.

Reviewer_N7JVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents OmniVaria, a blind omnidirectional video super-resolution framework that aims to account for the interaction between unknown image degradation and ERP projection geometry. Its Probabilistic Spherical Degradation Conditioning (PSDC) module infers geometry-conditioned local degradation representations and uses them to modulate reconstruction features. The method also introduces Implicit Omnidirectional Temporal Aggregation (IOTA), which combines current-frame features with recurrent hidden states without explicit optical-flow estimation or feature warping. Experiments on three omnidirectional video benchmarks show competitive results, including generalization to blur levels outside the training range.

Critical strengths
  1. The paper addresses a relevant and relatively underexplored problem. Extending blind video super-resolution to omnidirectional video requires accounting for both projection geometry and unknown degradation, which are not handled directly by conventional planar-video methods.
  2. The proposed design is coherent with the problem formulation. Latitude-wise degradation conditions provide a plausible way to model spatial variation in ERP, and flow-free temporal aggregation is a reasonable response to the difficulty of motion estimation under ERP distortion and blur.
  3. The experiments cover multiple datasets and degradation strengths. Testing at blur levels beyond the training range provides some evidence of robustness within the chosen degradation family. The ablation study compares several components, including local versus global conditions, geometry input, probabilistic versus deterministic conditions, variance gating, GCVIB, and IOTA. These experiments help assess the contribution of the overall design.
Critical weaknesses
  1. The degradation evaluation covers a limited family of synthetic degradations. The LR videos are generated using isotropic Gaussian blur and downsampling in the fisheye domain. Testing blur levels outside the training range is useful, but it evaluates extrapolation within the same degradation family. The paper does not establish whether the method generalizes to other plausible degradations or their combinations, such as spatially varying or motion blur, sensor noise, compression, and resampling artifacts. Since paired real-world HR/LR omnidirectional videos are difficult to obtain, a broader synthetic degradation study using the available HR references would be a practical way to assess robustness. The paper should also clarify this limitation when framing its claims about blind or real-world generalization.

  2. The claimed role of GCVIB is not directly established. The KL penalty encourages the posterior to remain close to a standard Gaussian prior, but this does not guarantee that the latent variables encode degradation rather than scene content or other reconstruction cues. The reported ablation shows that GCVIB can improve reconstruction, but the t-SNE visualization is insufficient to establish degradation disentanglement or content invariance. The performance gain from GCVIB is also modest in the reported averages: removing it reduces WS-PSNR by 0.06 dB on ODV360 and 0.19 dB on MiG Panorama.

3.The methodological novelty and motivation of IOTA could be better substantiated. The paper evaluates IOTA through ablations and comparisons with flow-warp- and DCN-based alignment, showing that it improves reconstruction quality in the tested settings. However, the analysis mainly reports aggregate metrics and does not explain why aggregating recurrent hidden states without explicit motion alignment is particularly suitable for the degradation considered in this work. For example, the paper does not analyze whether IOTA is more robust near ERP seams, under larger projected motion, or at latitudes where projection-induced deformation is stronger. Such analysis would help clarify the specific insight behind IOTA and distinguish its contribution from a general flow-free recurrent aggregation design.

More broadly, the paper’s methodological contribution appears to lie in combining established ideas for blind degradation conditioning, variational information bottleneck regularization, and recurrent temporal aggregation; the paper should more clearly identify what is technically new beyond this task-specific integration.

Questions for authors

Please refer to above weakness

3 comments

We sincerely thank the reviewer for the constructive feedback. Regarding your concerns, we hope the following responses address them adequately. We would also appreciate any further suggestions.

W.1 Generalization to Broader Degradations and Scope Clarification

Ans.: Thanks for your thorough comments which made us realize the importance of generalization. To this end, we conduct zero-shot tests on three unseen degradation families using ODV360, and further evaluate on FLOW360 [1], a synthetic benchmark featuring large motion and motion blur.

DegradationLevelOmniVariaFMA-NetSTDANOSRTw/o Prob.w/o GCVIB
Anisotropic Gaussian Blur(ηmaj, ηmin)WS-PSNR
(1.8, 0.6)26.1924.9925.4225.5024.8925.74
(2.4, 0.8)26.0924.8725.3825.2824.8125.63
(3.5, 0.8)25.6024.5124.8224.7324.3725.16
Gaussian Noiseη
525.4423.9823.1424.8724.0725.04
1024.0722.7122.0723.5822.6323.77
2021.5921.1720.1621.0720.2721.24
H.264 CompressionCRF
1825.6924.7724.6925.1624.3825.27
2325.2424.5024.5124.8224.4824.91
2824.4923.9724.0024.2523.9924.28
Large Motion & Motion BlurFLOW36033.6132.3230.7132.4432.2533.04

OmniVaria consistently outperforms the baselines, showing promising generalization capability toward unknown degradations. We will report these results and clarify the synthetic evaluation scope and real-world limitations in the revised manuscript.

W.2 The claimed role of GCVIB is not directly established.

Ans.: Thank you for this insightful comment, which helped us realize that our phrasing regarding GCVIB's contribution to scene content suppression was inappropriate. We have revised the manuscript to align our description with the mathematical and experimental support.

Since PSDC is optimized for HR reconstruction, its conditions are not intended to represent degradation identities exclusively, but to provide geometry-conditioned, reconstruction-relevant cues for local HR reconstruction. Thus, OmniVaria exhibits partial overlap among neighboring low-latitude conditions while maintaining distinct representations for polar regions in t-SNE. Without GCVIB, the conditions form compact clusters corresponding to pairs of symmetric latitudes. More importantly, the zero-shot results in W.1 demonstrate a pronounced contribution of GCVIB under unseen degradation distributions.

Thus, GCVIB applies the variational information bottleneck to geometry-conditioned, latitude-structured probabilistic representations. Its KL regularization constrains the information retained from LR observations given the ERP geometry prior, while reconstruction supervision encourages preserving complementary information useful for HR recovery. This information-capacity constraint may alleviate over-specialization to training degradation statistics, improving generalization.

W.3 Analysis of IOTA

Ans.: Thanks for your valuable comments. Following your suggestion, we further compare IOTA with flow warping and DCN using latitude-wise PSNR, ERP-seam Discontinuity Score[2] (DS), and efficiency.

PSNR by [0°, 30°] [30°, 60°] [60°, 90°]Seam DS↓Param. (M)GMACsLatency (ms/frame)
IOTA25.29/27.51/28.460.49050.101122.517.64
Flow Warp25.09/27.28/28.190.50651.441909.4711.22
DCN25.10/27.34/28.330.49332.2710870.1249.21
FLOW360PSNR by [0°, 30°] [30°, 60°] [60°, 90°]
IOTA32.46/35.54/38.83
Flow Warp32.29/35.33/38.57
DCN32.33/35.46/38.85

IOTA gets rid of the computational cost and uncertainty introduced by flow estimation and achieves higher performance on ODV360, especially at higher latitudes. Only when encountering large motion does IOTA perform as a runner-up due to its limited receptive field. Thus, IOTA could be a robust, efficient temporal aggregation module suitable for reconstructing ERP content.

W.4 Technical contributions beyond task-specific integration.

Ans.: Thanks for your constructive comments. OmniVaria follows a unified geometry-driven design for blind ODVSR. Specifically, fisheye degradation and ERP reprojection induce spatially varying degradations, motivating PSDC to infer latitude-structured Gaussian posteriors jointly from LR observations and spherical geometry to guide spatially varying feature modulation, while GCVIB regularizes their information capacity to promote reconstruction. Meanwhile, projection-dependent motion complicates temporal correspondence, motivating IOTA to implicitly aggregate temporal information. Thus, spherical geometry serves as a common design rationale linking degradation inference, spatial adaptation, and temporal aggregation. Our component ablations and additional experiments support these geometry-motivated designs.

[1] Learning Omnidirectional Flow in 360° Video via Siamese Representation, ECCV 2022.

[2] Geometry Fidelity for Spherical Images, ECCV 2024.

Area_Chair_4uxJOpenReview ↗

Reminder for Discussion

Dear Reviewer,

The authors have answered your questions. Please have a check and give your feedback asap.

Thanks for your contribution to ICLR 2027.

Your AC

Thank the authors for the additional experiments and clarifications and sorry for late reply. Given that this field lacks real-world degradation data, the first experiment (unseen-degradation testing) does demonstrate a certain extrapolation capability, and removing GCVIB degrades performance under unseen degradations. This evidence is relatively strong and essentially addresses my first two concerns.

Regarding my third concern (the analysis of IOTA), I thank the authors for providing the latitude-wise PSNR, ERP-seam DS, and efficiency comparison data. These data do respond to the analytical dimensions I previously requested, and I acknowledge IOTA's advantages in efficiency and at higher latitudes. However, I still fail to understand mechanistically why implicit recurrent aggregation (without explicit motion alignment) is particularly suitable for the degradation considered in this work—and especially why IOTA performs better in higher-latitude regions where projection-induced deformation is stronger. The current data show where it performs better, but do not yet explain why it performs better there, nor the causal link between this advantage and the degradation characteristics considered in this paper. I would appreciate further clarification of this mechanism from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.