Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

SO3UFormer claims that a spherical Transformer can segment panoramas accurately under arbitrary SO(3) rotations—retaining 70.67 mIoU where a strong baseline collapses to 25.26—by removing absolute latitude cues and using gauge-aware, quadra

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 20:34 UTC pith:BXBLOSNP

load-bearing objection The rotation-robustness result is real and the ablation is clean, but the abstract overstates what the paper actually shows—the missing Matterport3D/depth experiments and the 'OOD' label for a transformation the regularizer already trains on need fixing. the 4 major comments →

arxiv 2602.22867 v2 pith:BXBLOSNP submitted 2026-02-26 cs.CV

SO3UFormer: Learning Intrinsic Spherical Features for Rotation-Robust Panoramic Dense Prediction

Qinfeng Zhu , Yunxi Jiang , Lei Fan This is my paper
classification cs.CV
keywords panoramic semantic segmentationrotation robustnessspherical transformerSO(3) equivariancegauge-aware attentionquadrature-consistent attentionicosahedral discretizationPose35 benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that current panoramic segmentation models are fragile to 3D camera rotation because they rely on absolute latitude encoding and non-uniform spherical sampling. It proposes SO3UFormer, which removes gravity-dependent positional cues, corrects attention for sphere sampling density, and uses local tangent-plane angles for relative position. On the new Pose35 benchmark and a full SO(3) stress test, the method keeps accuracy nearly constant while the leading SphereUFormer baseline collapses from 67.53 to 25.26 mIoU. A training-time logit-space consistency regularizer further reduces discretization artifacts. If correct, this largely closes the rotation-robustness gap for panoramic dense prediction.

Core claim

The central claim is that rotation robustness in panoramic dense prediction can be achieved by replacing global coordinate shortcuts with intrinsic spherical geometry. Concretely, SO3UFormer removes absolute latitude positional encoding (the main source of gravity bias), weights attention logits by spherical area (log ω_j) so aggregation matches the manifold's non-uniform sampling, and defines relative positional bias from tangent-plane projected angles averaged over six in-plane gauge rotations. Geometry-consistent up/down sampling, plus a training-only SO(3)-consistency regularizer based on index-based resampling, closes the remaining discretization gap. The result is that mIoU under arbit

What carries the argument

The load-bearing components are: (1) removal of absolute latitude encoding; (2) quadrature-consistent attention, where each neighbor's logit is offset by the log of its spherical area weight to correct sampling density; (3) a gauge-pooled Fourier relative positional bias built from tangent-plane angles around each query, pooled over six in-plane rotations to cancel global reference; (4) area-weighted downsampling and geodesic-kernel upsampling; and (5) a training-time logit-space MSE regularizer that enforces consistency under spherical index resampling. Together these make the learned features intrinsic—defined by local spherical geometry rather than the chosen coordinate frame.

Load-bearing premise

The central premise is that the full-SO(3) stress test is a fair apples-to-apples comparison: because SO3UFormer's regularizer samples a random 3D rotation at every training iteration, the test rotations fall inside its training distribution, whereas baselines only saw yaw/flip augmentation; if a baseline were given the same rotation sampling, the headline margin could shrink.

What would settle it

Retrain the strongest baseline (SphereUFormer) with exactly the same per-iteration uniform random rotation resampling used by SO3UFormer, without the proposed attention changes. If its full-SO(3) mIoU rises from 25.26 to near 70, the architectural contributions are not the main driver of robustness.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Panoramic segmentation models can embed rotation robustness architecturally rather than through full equivariant group convolutions, making the design practical for existing transformer backbones.
  • Dense prediction on handheld and aerial platforms with rotational jitter becomes feasible without gravity-alignment preprocessing or test-time pose correction.
  • Pose35 and the full-SO(3) stress test provide a standardized protocol for measuring orientation robustness beyond yaw-only benchmarks.
  • The same geometric recipe transfers to panoramic depth estimation, where the paper reports similar stability under rotation.
  • Removing absolute latitude encoding is the single most decisive change, lifting SO(3) mIoU from 25.26 to 64.66, so future spherical models should avoid global axis biases.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The residual gap between 70.67 and 72.03 mIoU suggests perfect equivariance is not reached; a continuous gauge average over all in-plane rotations, instead of six discrete ones, might close that gap further.
  • Because the training-time regularizer samples a random rotation every iteration, the stress-test rotations fall inside SO3UFormer's training distribution; a baseline trained with the same resampling but without the geometric attention would isolate whether the architecture or the augmentation drives the gain.
  • The same principle—local tangent-plane geometry over global coordinates—could extend to other spherical tasks such as object detection and 6-DoF tracking under large viewpoint changes.
  • Real attitude-variation datasets (drone or handheld captures with tracked pose) would provide a more direct validation than synthetic rotations, as the paper itself notes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SO3UFormer, a spherical U-shaped Transformer for panoramic semantic segmentation that is designed to be robust to arbitrary 3D rotations. The method removes absolute latitude encoding, introduces quadrature-consistent attention with area weights, uses a gauge-pooled Fourier relative positional bias, applies geometry-consistent down/up sampling, and adds a logit-space SO(3)-consistency regularizer used only during training. The authors introduce Pose35, a rotation-perturbed variant of Stanford2D3D, and evaluate under a full-SO(3) stress test. They report that SO3UFormer achieves 72.03 mIoU on Pose35 and 70.67 mIoU under the stress test, whereas SphereUFormer drops from 67.53 to 25.26 mIoU. The abstract also claims robustness on Matterport3D segmentation and panoramic depth estimation, but these results do not appear in the body.

Significance. If the claims are supported, this is a useful contribution to a real problem: panoramic models trained under gravity-aligned assumptions are known to fail under roll/pitch changes. The paper's main architectural idea—removing global latitude cues and using local tangent-plane geometry—is reasonable and the ablation suggests that the architecture alone, even without the regularizer, yields a large robustness gain (69.90 vs. 25.26 mIoU under the SO(3) stress test). The Pose35 benchmark and the explicit stress-test protocol are valuable additions. However, the central quantitative comparison is weakened by a train/test transformation mismatch for the regularized model, and the abstract promises experiments that are not in the manuscript. The paper is therefore not yet ready for acceptance, but the core direction is credible and the issues are addressable.

major comments (4)
  1. [Sec. 3.5, Eq. (18)-(20); Sec. 4.1] The SO(3) stress test is not a true out-of-distribution test for the final SO3UFormer model. The regularizer in Eq. (18) uses the same nearest-neighbor index mappings idxproj(g) and idximg(g) that the stress test applies to spherical-token models (Sec. 4.1: 'nearest-neighbor index mapping over icosphere normals'). Since the regularizer samples a uniform random rotation every iteration, the full-SO(3) test rotations are inside the training distribution of the regularizer. Baselines, by contrast, are evaluated with inverse-map ERP remapping, a transformation they have not seen during training. This confounds the headline margin in Table 2. The ablation in Table 1 shows that the architecture alone still gives 69.90, so the core claim is plausible, but the comparison should be made fair by training baselines with the same rotation augmentation/regularizer, or by evaluating SO3UFormer with a
  2. [Abstract; Sec. 4] The abstract claims that 'on a second real-world dataset (Matterport3D) for segmentation and on panoramic depth estimation, SO3UFormer remains essentially rotation-invariant while the gravity-anchored baseline again loses most of its accuracy.' No Matterport3D or depth experiments appear anywhere in the body; Section 4 reports only Pose35 semantic segmentation (Tables 1-2 and Figure 3). The depth claim is also contradicted by the conclusion, which lists depth as future work. These claims must either be substantiated with experiments or removed from the abstract and contributions.
  3. [Tables 1 and 2; Sec. 4.1] No error bars, confidence intervals, or repeated-seed results are reported. The stress test uses 10 rotations repeated 3 times, but only the mean is given. Since the central comparison relies on a single run of each model and the baseline collapse is very large, the report should include per-seed and per-rotation variability. Otherwise it is difficult to assess whether the 2-3 mIoU differences between ablations are meaningful, and whether the baseline's low SO(3) mIoU is stable.
  4. [Title/Abstract; Sec. 3.5] The paper's title and abstract promise 'intrinsic spherical features' and near-rotation-invariance, but the method is not equivariant by construction and the regularizer is explicitly based on a discrete nearest-neighbor reindexing approximation (Sec. 3.5: 'rather than an exact continuous SO(3) action'). The wording 'largely decoupled from the underlying coordinate frame' is stronger than what is demonstrated. The paper should more carefully distinguish approximate discrete rotation robustness from true continuous SO(3) invariance, and avoid overclaiming in the abstract and conclusion.
minor comments (5)
  1. [Sec. 4.2, Table 1] The table caption repeats 'Best results are in bold.' Remove the duplicate.
  2. [Sec. 3.5, Eq. (18)] The notation x0[idxproj(g)] is ambiguous: it should state explicitly whether this means gathering input tokens at the index-mapped node positions and whether the same mapping is applied to the output logits. This is important because the regularizer's target is z[idximg(g)] with stopgrad, and the exact indexing convention affects reproducibility.
  3. [Sec. 3.3, Eq. (13)] The values of the Fourier truncation order M, the number of anchors F, and the radial bin count B are not reported in the experiments. These are free parameters and should be listed with the other training details.
  4. [Sec. 3.4, Eq. (16)] The upsampling bandwidth sigma is called 'fixed', but its value is not given. Please specify sigma and the icosphere ranks used in the hierarchy.
  5. [Sec. 4.1] Pose35 is introduced as a 'dataset variant', but it is created by deterministically rotating each sample. The paper should state explicitly whether the validation set is also rotated and whether the same fixed rotations are used for all methods. This is essential for reproducing the stress test.

Circularity Check

1 steps flagged

SO(3) stress-test transform matches the regularizer's training-time index mapping, so the final 70.67 mIoU is partly in-distribution; architectural ablation still gives independent support.

specific steps
  1. other [Sec. 3.5 Eq. (18)-(20); Sec. 4.1 'Training protocol' and 'SO(3) OOD stress test']
    "When the SO(3)-consistency regularizer is enabled, we sample one random 3D rotation per iteration (uniform quaternion), apply the corresponding index-based resampling, and add the logit-space MSE penalty (Sec. 3.5). ... To ensure a fair comparison across representations, we use the same set of 3D rotations for all methods, while applying them in each method’s native input domain—nearest-neighbor index mapping over icosphere normals for spherical-token models, and inverse-map ERP remapping for ERP-based baselines."

    Eq. (18)-(19) train the model so that zrot = ftok(x0[idxproj(g)]) matches ztgt = stopgrad(z[idximg(g)]), i.e. predictions are made consistent under a discrete nearest-neighbor index-based spherical resampling. The full-SO(3) 'OOD' stress test applies rotations to spherical-token models using exactly 'nearest-neighbor index mapping over icosphere normals'. Because training rotations are uniform quaternions, the test rotations are inside the regularizer's training distribution, not out-of-distribution for the final model. Thus the reported 70.67 SO(3) mIoU is partly obtained by training on the same transformation that is later presented as an independent stress test; the comparison with ERP baselines, which receive a different transform they never trained on, is confounded. This is not a mat

full rationale

Most of this paper is an architecture-and-benchmark paper with no formal derivation chain, so there is no candidate for definitional circularity. The core geometric claims are independently ablated: removing absolute latitude encoding lifts SO(3) mIoU from 25.26 to 64.66, and the full architecture without the SO(3)-consistency regularizer reaches 69.90. These results do not reduce to the inputs by construction. The one substantive circularity concern is the benchmark-level overlap between the regularizer and the stress test. Eq. (18)-(20) use idxproj(g)/idximg(g), discrete nearest-neighbor index mappings, to enforce logit consistency under uniform random rotations; Sec. 4.1 evaluates spherical-token models under 'nearest-neighbor index mapping over icosphere normals', the same operation, with rotations covering all of SO(3). Consequently, the final model's full-SO(3) result is partly in-distribution with respect to the regularizer, and the paper's 'out-of-distribution' framing is inaccurate for the regularized model. The paper even concedes 'Higher-fidelity spherical resampling for training-time consistency remains an open avenue,' acknowledging the discretization limitation. Because the architectural ablation still gives large, independent robustness gains, this is a partial confound rather than a fully circular derivation; hence score 4 rather than 6 or higher.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 1 invented entities

The empirical claims rest on multiple unspecified design constants (F, M, B, sigma, lambda, rotation ranges) and on the validity of discrete icosphere resampling as a proxy for continuous SO(3). No new physical entities are introduced; Pose35 is an evaluation artifact rather than a theoretical entity.

free parameters (8)
  • SO(3)-consistency regularizer weight lambda = 0.05
    Fixed to 0.05 in the final ablation (Table 1); sensitivity not analyzed.
  • Number of gauge anchors F = not reported
    Anchor count in the tangent_max rule for gauge-pooled bias (Sec 3.3); not swept.
  • Fourier truncation order M = not reported
    Defines angular modes in Eq. (13); affects bias expressiveness and computational cost.
  • Radial bins B = not reported
    Tabulated distance-dependent coefficients are interpolated over B bins (Eq. 11).
  • Upsampling bandwidth sigma = not reported
    Fixed bandwidth in the geodesic upsampling kernel (Eq. 16); value not given.
  • Attention scale upper clamp = 100
    Clamp in Eq. (4) prevents extremely sharp attention; chosen by hand.
  • Pose35 max rotation magnitude = 35 degrees
    Benchmark perturbation range; arbitrary choice for the stress protocol.
  • SO(3) stress-test rotation count and repeats = 10 rotations, 3 repeats
    Number of Euler rotations per image and repeats; no standard deviation reported.
axioms (5)
  • domain assumption Icosahedral vertex grid with area weights is a sufficient quadrature for spherical attention.
    Sec 3.1; all attention is local on the icosphere; the quadrature correction (Eq. 5) assumes this discretization approximates integration on S^2.
  • ad hoc to paper Nearest-neighbor index resampling is a benign approximation to the SO(3) action.
    Sec 3.5 calls it 'a discrete nearest-neighbor reindexing/resampling approximation ... rather than an exact continuous SO(3) action'. The robustness numbers depend on this approximation being unbiased.
  • ad hoc to paper Removing absolute latitude encoding does not discard task-critical information.
    This is the first design pillar (Sec 3.2). Table 1 supports it empirically on this dataset, but it need not generalize to all spherical dense-prediction tasks.
  • ad hoc to paper Self-distillation in logit space is a valid training signal for rotation consistency.
    Eq. (18)-(20) use the model's own stopped-gradient predictions as targets; consistency is enforced, not independently measured.
  • domain assumption Stanford2D3D with 39 validation panoramas is representative for evaluating rotation robustness.
    Sec 4.1; the validation set is small and no confidence intervals are given.
invented entities (1)
  • Pose35 dataset variant no independent evidence
    purpose: Evaluation benchmark created by applying deterministic per-sample rotations up to ±35° to Stanford2D3D panoramas.
    Introduced by the authors (Sec 4.1) with no external validation of whether the perturbations represent real deployment conditions.

pith-pipeline@v1.3.0-alltime-deepseek · 12507 in / 15474 out tokens · 149925 ms · 2026-08-02T20:34:03.266062+00:00 · methodology

0 comments
read the original abstract

Panoramic dense-prediction models, spanning semantic segmentation and depth estimation, are typically trained under a strict gravity-aligned assumption. Real-world captures, however, routinely violate it: handheld devices jitter and aerial platforms change attitude, so the camera is rarely upright. Under such 3D reorientation, standard spherical Transformers overfit global latitude cues and collapse. We introduce SO3UFormer, an architecture that learns intrinsic spherical features largely decoupled from the underlying coordinate frame, through three geometric components: (1) removing absolute latitude encoding, which breaks the dependence on the gravity axis; (2) quadrature-consistent spherical attention, which corrects for non-uniform sampling density; and (3) a gauge-aware relative positional bias built from local tangent-plane angles rather than global axes. A logit-space \emph{SO(3)}-consistency regularizer, used only during training, further suppresses residual discretization effects. To benchmark robustness, we introduce Pose35, a variant of Stanford2D3D perturbed by random rotations within $\pm 35^\circ$, and evaluate under a full, arbitrary \emph{SO(3)} stress test. There, the baseline SphereUFormer collapses from 67.53 \emph{mIoU} on Pose35 to 25.26 under the full \emph{SO(3)} test, whereas SO3UFormer reaches 72.03 on Pose35 and retains 70.67 under the same test. Similarly, on a second real-world dataset (Matterport3D) for segmentation and on panoramic depth estimation, SO3UFormer remains essentially rotation-invariant while the gravity-anchored baseline again loses most of its accuracy. Code and models are available at https://github.com/zhuqinfeng1999/SO3UFormer.

Figures

Figures reproduced from arXiv: 2602.22867 by Lei Fan, Qinfeng Zhu, Yunxi Jiang.

Figure 1
Figure 1. Figure 1: Breaking the Gravity Lock: Rotation Robustness in Panoramic Seg [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SO3UFormer overview. A U-shaped spherical Transformer with gauge￾aware, quadrature-consistent local attention, geometry-consistent down/up sampling, and an SO(3)-consistency regularizer via spherical index-based resampling. Notation. At an internal spherical-token scale s, let \protect \mathbf {X}^{(s)}\in \mathbb {R}^{L_s\times D} denote node features, where L_s is the number of nodes at that rank. When t… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison under the out-of-distribution [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Panoramic Scene Understanding: A Survey from Distortion-Aware Engineering to Sphere-Native Modeling

    cs.CV 2026-06 conditional novelty 5.0

    A survey diagnosing panoramic scene understanding as a field that converged on compatibility-preserving geometric adaptation rather than sphere-native modeling, while its evaluation protocols systematically fail to me...

  2. Panoramic Scene Understanding: A Survey from Distortion-Aware Engineering to Sphere-Native Modeling

    cs.CV 2026-06 unverdicted novelty 3.0

    Survey organizing panoramic scene analysis literature by architectural design and training paradigm, identifying the absence of methods achieving both strict spherical equivariance and full reuse of perspective-pretra...

Reference graph

Works this paper leans on

58 extracted references · 7 linked inside Pith · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2205.10468 (2022)

    Ai, H., Cao, Z., Zhu, J., Bai, H., Chen, Y., Wang, L.: Deep learning for omnidi- rectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468 (2022)

    Pith/arXiv arXiv 2022

  2. [2]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Ai, H., Wang, L.: Elite360d: Towards efficient 360 depth estimation via semantic- and distance-aware bi-projection fusion. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 9926–9935 (2024)

    2024

  3. [3]

    Robotica42(7), 2089–2107 (2024)

    Arif, M.A., Zhu, A., Mao, H., Tu, Y.: Panoramic visual system for spherical mobile robots. Robotica42(7), 2089–2107 (2024)

    2089

  4. [4]

    arXiv preprint arXiv:1702.01105 (2017)

    Armeni, I., Sax, S., Zamir, A.R., Savarese, S.: Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105 (2017)

    Pith/arXiv arXiv 2017

  5. [5]

    In: Proceedings of the Computer Vision and Pattern Recognition Confer- ence

    Benny, Y., Wolf, L.: Sphereuformer: A u-shaped transformer for spherical 360 per- ception. In: Proceedings of the Computer Vision and Pattern Recognition Confer- ence. pp. 940–950 (2025)

    2025

  6. [6]

    In: 2025 IEEE International Conference on Image Processing (ICIP)

    Bergmann, M.A., Stringhini, R.M., da Silveira, T.L., Jung, C.R.: Anchor-based gravity alignment for panoramas. In: 2025 IEEE International Conference on Image Processing (ICIP). pp. 2187–2192. IEEE (2025)

    2025

  7. [7]

    arXiv preprint arXiv:2505.11157 (2025)

    Bonev, B., Rietmann, M., Paris, A., Carpentieri, A., Kurth, T.: Attention on the sphere. arXiv preprint arXiv:2505.11157 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    IEEE Signal Processing Magazine 34(4), 18–42 (2017)

    Bronstein, M.M., Bruna, J., LeCun, Y., Szlam, A., Vandergheynst, P.: Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine 34(4), 18–42 (2017)

    2017

  9. [9]

    IEEE Transactions on robotics32(6), 1309– 1332 (2017)

    Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics32(6), 1309– 1332 (2017)

    2017

  10. [10]

    In: 2024 4th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICEC- CME)

    Candela, F., Morabito, A.F., Morabito, F.C.: Object recognition in high-resolution 360 panoramic images using spherical grids. In: 2024 4th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICEC- CME). pp. 1–6. IEEE (2024)

    2024

  11. [11]

    Advances in Neural Information Processing Systems37, 26355–26376 (2024)

    Cao Dinh, D., Kim, S.J., Cho, K.: Geometric exploitation for indoor panoramic semantic segmentation. Advances in Neural Information Processing Systems37, 26355–26376 (2024)

    2024

  12. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Carlsson, O., Gerken, J.E., Linander, H., Spieß, H., Ohlsson, F., Petersson, C., Persson, D.: Heal-swin: A vision transformer on the sphere. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6067– 6077 (2024)

    2024

  13. [13]

    JournalofKingSaudUniversity-ComputerandInformationSciences35(5),101546 (2023)

    Chen, H., Zhao, J.: 3d mesh classification and panoramic image segmentation us- ing spherical vector networks with rotation-equivariant self-attention mechanism. JournalofKingSaudUniversity-ComputerandInformationSciences35(5),101546 (2023)

    2023

  14. [14]

    In: Proceedings of the European conference on computer vision (ECCV)

    Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801–818 (2018) 16 Q. Zhu et al

    2018

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1290–1299 (2022)

    2022

  16. [16]

    In: International conference on Machine learning

    Cohen, T., Weiler, M., Kicanaoglu, B., Welling, M.: Gauge equivariant convolu- tional networks and the icosahedral cnn. In: International conference on Machine learning. pp. 1321–1330. PMLR (2019)

    2019

  17. [17]

    In: Interna- tional conference on machine learning

    Cohen, T., Welling, M.: Group equivariant convolutional networks. In: Interna- tional conference on machine learning. pp. 2990–2999. PMLR (2016)

    2016

  18. [18]

    arXiv preprint arXiv:1801.10130 (2018)

    Cohen, T.S., Geiger, M., Köhler, J., Welling, M.: Spherical cnns. arXiv preprint arXiv:1801.10130 (2018)

    Pith/arXiv arXiv 2018

  19. [19]

    In: Proceedings of the European conference on computer vision (ECCV)

    Coors, B., Condurache, A.P., Geiger, A.: Spherenet: Learning spherical represen- tations for detection and classification in omnidirectional images. In: Proceedings of the European conference on computer vision (ECCV). pp. 518–533 (2018)

    2018

  20. [20]

    In: 2009 IEEE conference on computer vision and pattern recognition

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE conference on computer vision and pattern recognition. pp. 248–255. Ieee (2009)

    2009

  21. [21]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Eder,M.,Shvets,M.,Lim,J.,Frahm,J.M.:Tangentimagesformitigatingspherical distortion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12426–12434 (2020)

    2020

  23. [23]

    arXiv preprint arXiv:2111.10250 (2021)

    Elharrouss, O., Al-Maadeed, S., Subramanian, N., Ottakath, N., Almaadeed, N., Himeur, Y.: Panoptic segmentation: A review. arXiv preprint arXiv:2111.10250 (2021)

    Pith/arXiv arXiv 2021

  24. [24]

    Advances in neural information pro- cessing systems33, 1970–1981 (2020)

    Fuchs, F., Worrall, D., Fischer, V., Welling, M.: Se (3)-transformers: 3d roto- translation equivariant attention networks. Advances in neural information pro- cessing systems33, 1970–1981 (2020)

    1970

  25. [25]

    IEEE Transactions on Instrumentation and Measurement71, 1–34 (2022)

    Gao, S., Yang, K., Shi, H., Wang, K., Bai, J.: Review on panoramic imaging and its applications in scene understanding. IEEE Transactions on Instrumentation and Measurement71, 1–34 (2022)

    2022

  26. [26]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Guttikonda, S., Rambach, J.: Single frame semantic segmentation using multi- modal spherical images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3222–3231 (2024)

    2024

  27. [27]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  28. [28]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference

    Jung, D., Choi, J., Lee, Y., Jeong, S., Lee, T., Manocha, D., Yeon, S.: Edm: Equirectangular projection-oriented dense kernelized feature matching. In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference. pp. 6337– 6347 (2025)

    2025

  29. [29]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing17, 20026–20040 (2024)

    Liu, A., Meng, L., Xiao, L.: Visual rotated position encoding transformer for re- mote sensing image captioning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing17, 20026–20040 (2024)

    2024

  30. [30]

    In: International Conference on Image, Signal Processing, and Pattern Recognition (ISPP 2023)

    Liu, M., Ding, Y., Xie, Y.: Uav-based panoramic scene recognition utilizing discrete spherical image features. In: International Conference on Image, Signal Processing, and Pattern Recognition (ISPP 2023). vol. 12707, pp. 619–625. SPIE (2023) SO3UFormer 17

    2023

  31. [31]

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

    2021

  32. [32]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3431–3440 (2015)

    2015

  33. [33]

    Computational Mechanics42(6), 907–919 (2008)

    Mäkinen, J.: Rotation manifold so (3) and its tangential vectors. Computational Mechanics42(6), 907–919 (2008)

    2008

  34. [34]

    Astronomy and Computing27, 130–146 (2019)

    Perraudin, N., Defferrard, M., Kacprzak, T., Sgier, R.: Deepsphere: Efficient spher- ical convolutional neural network with healpix sampling for cosmological applica- tions. Astronomy and Computing27, 130–146 (2019)

    2019

  35. [35]

    ACM Transactions on Graphics (TOG)40(6), 1–12 (2021)

    Pintore, G., Almansa, E., Agus, M., Gobbetti, E.: Deep3dlayout: 3d reconstruc- tion of an indoor layout from a spherical panoramic image. ACM Transactions on Graphics (TOG)40(6), 1–12 (2021)

    2021

  36. [36]

    IEEE transactions on robotics34(4), 1004–1020 (2018)

    Qin, T., Li, P., Shen, S.: Vins-mono: A robust and versatile monocular visual- inertial state estimator. IEEE transactions on robotics34(4), 1004–1020 (2018)

    2018

  37. [37]

    In: International Conference on Medical image computing and computer-assisted intervention

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

    2015

  38. [38]

    Multimedia Tools and Applications84(35), 44705–44726 (2025)

    Sahin, C., Ergun, B., Bilucan, F.: Non-uniform image coordinate system defini- tion for panoramic imaging using a hybrid ai algorithm. Multimedia Tools and Applications84(35), 44705–44726 (2025)

    2025

  39. [39]

    In: International conference on machine learning

    Satorras, V.G., Hoogeboom, E., Welling, M.: E (n) equivariant graph neural net- works. In: International conference on machine learning. pp. 9323–9332. PMLR (2021)

    2021

  40. [40]

    In: European Conference on Com- puter Vision

    Shen, Z., Lin, C., Liao, K., Nie, L., Zheng, Z., Zhao, Y.: Panoformer: panorama transformer for indoor 360◦ depth estimation. In: European Conference on Com- puter Vision. pp. 195–211. Springer (2022)

    2022

  41. [41]

    In: 2013 IEEE International Conference on Image Processing

    Shidanshidi, H., Safaei, F., Zamani-Farahani, A., Li, W.: Non-uniform sampling of plenoptic signal based on the scene complexity variations for a free viewpoint video system. In: 2013 IEEE International Conference on Image Processing. pp. 3147–3151. IEEE (2013)

    2013

  42. [42]

    Advances in neural information processing systems30(2017)

    Su, Y.C., Grauman, K.: Learning spherical convolution for fast features from 360 imagery. Advances in neural information processing systems30(2017)

    2017

  43. [43]

    In: Geometric Deep Learning in Medical Image Analysis (2022)

    Suliman, M.A., Williams, L.Z.J., Fawaz, A., Robinson, E.C.: Geomorph: Geometric deep learning for cortical surface registration. In: Geometric Deep Learning in Medical Image Analysis (2022)

    2022

  44. [44]

    In: 2021 European Con- ference on Mobile Robots (ECMR)

    Sun, L., Wang, J., Yang, K., Wu, K., Zhou, X., Wang, K., Bai, J.: Aerial-pass: Panoramic annular scene segmentation in drone videos. In: 2021 European Con- ference on Mobile Robots (ECMR). pp. 1–6. IEEE (2021)

    2021

  45. [45]

    arXiv preprint arXiv:1802.08219 (2018)

    Thomas, N., Smidt, T., Kearnes, S., Yang, L., Li, L., Kohlhoff, K., Riley, P.: Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219 (2018)

    Pith/arXiv arXiv 2018

  46. [46]

    Virtual Reality & Intelligent Hardware4(3), 223–246 (2022)

    Ullah, H., Afzal, S., Khan, I.U.: Perceptual quality assessment of panoramic stitched contents for immersive applications: a prospective survey. Virtual Reality & Intelligent Hardware4(3), 223–246 (2022)

    2022

  47. [47]

    Advances in neural information pro- cessing systems30(2017) 18 Q

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017) 18 Q. Zhu et al

    2017

  48. [48]

    Advances in neural information processing systems32(2019)

    Weiler, M., Cesa, G.: General e (2)-equivariant steerable cnns. Advances in neural information processing systems32(2019)

    2019

  49. [49]

    IEEE Transactions on Circuits and Systems for Video Technology (2025)

    Wu,X.,Liu,S.,Chang,R.,Fu,Q.,He,W.:Digitalvideostabilizationmethodbased on jitter analysis of flapping-wing flying robots. IEEE Transactions on Circuits and Systems for Video Technology (2025)

    2025

  50. [50]

    Advances in neural information processing systems34, 12077–12090 (2021)

    Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems34, 12077–12090 (2021)

    2021

  51. [51]

    IEEE Trans- actions on Intelligent Transportation Systems (2025)

    Xu, H., Liu, X., Liu, Y., Song, X.: Motion estimation and reconstruction for 360◦ panoramic images based on variant goldberg polyhedral projection. IEEE Trans- actions on Intelligent Transportation Systems (2025)

    2025

  52. [52]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yoon, Y., Chung, I., Wang, L., Yoon, K.J.: Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5677–5686 (2022)

    2022

  53. [53]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhang, C., Wu, Q., Gambardella, C.C., Huang, X., Phung, D., Ouyang, W., Cai, J.: Taming stable diffusion for text to 360 panorama image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6347–6357 (2024)

    2024

  54. [54]

    Computer Science Review56, 100721 (2025)

    Zhang, J., Chen, X., Yang, B., Guan, Q., Chen, Q., Chen, J., Wu, Q., Xie, Y., Xia, Y.: Advances in attention mechanisms for medical image segmentation. Computer Science Review56, 100721 (2025)

    2025

  55. [55]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2881–2890 (2017)

    2017

  56. [56]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing17, 10454–10467 (2024)

    Zheng, H., Chang, Z., Li, Y., Zhu, J., Wang, W., Yang, Q., Xie, C., Zhang, J., Liu, J.: An efficient and fast image mosaic approach for highway panoramic uav im- ages. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing17, 10454–10467 (2024)

    2024

  57. [57]

    In: European Conference on Computer Vision

    Zheng, J., Liu, R., Chen, Y., Peng, K., Wu, C., Yang, K., Zhang, J., Stiefelhagen, R.: Open panoramic segmentation. In: European Conference on Computer Vision. pp. 164–182. Springer (2024)

    2024

  58. [58]

    Neural Networks p

    Zhu, Q., Jiang, Y., Fan, L.: Classwise-crf: Category-specific fusion for enhanced se- mantic segmentation of remote sensing imagery. Neural Networks p. 108485 (2025)

    2025