Pith. sign in

REVIEW 4 major objections 6 minor 63 references

SphereFusion: Efficient Panorama Depth Estimation via Gated Fusion

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read SphereFusion claims that fusing equirectangular and spherical features with learned gates produces panorama depth accuracy competitive with state-of-the-art methods while running at about 60 frames per second on 512x1024 images.

desk verdict Solid fusion method for panorama depth, but the "fastest inference" claim rests on uncontrolled baseline timings and needs a same-machine re-benchmark. read the letter →

arxiv 2502.05859 v1 pith:WN66TYLJ submitted 2025-02-09 cs.CV

Qingsong Yan , Qiang Wang , Kaiyong Zhao , Jie Chen , Bo Li , Xiaowen Chu , Fei Deng This is my paper · ORCID
classification cs.CV
keywords panoramadepthestimation360-degreeimagessphericalprojectionequirectangulargatedfusionmeshconvolutionreal-timeinferenceomnidirectionalvision
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

SphereFusion is an end-to-end method for estimating depth from a single 360-degree panorama. Its central claim is that no single projection is enough: equirectangular images give rich texture but distort at the poles, while spherical meshes avoid distortion and discontinuity but lose detail. The paper fuses features from both projections in the spherical domain using a gate that learns which features to trust, and it claims this reaches accuracy competitive with state-of-the-art methods on three public datasets while running at about 60 frames per second (17 ms per 512x1024 image) on an RTX 3090. The payoff, if true, is real-time omnidirectional depth for robot sensing and driving with a lightweight network.

What carries the argument

The load-bearing mechanism is GateFuse, a fusion module that combines spherical features $F_{sp}$ and equirectangular features $F_{eq}$ in the spherical domain using a reset gate $r$ and a forget gate $z$, so the network learns which projection to trust at each location rather than assuming one is better. Supporting it are the E2S/S2E projection mappings between equirectangular pixels and spherical triangle centers, mesh convolution and pooling built on face-adjacent-face connectivity from the icosahedron mesh, and a cache that stores the adjacency structure of active meshes to avoid recomputing it per layer.

What would settle it

Run the compared baselines on the same GPU with the same 512x1024 input and the same software stack; if any of them matches or beats 17 ms per image, the paper's headline efficiency claim fails.

Watch

Extended reading notes

Core claim

The paper's discovery is that depth estimation performed entirely on an icosahedral spherical mesh can be made both accurate and fast if the mesh features are augmented with equirectangular features through a gated fusion module. SphereFusion extracts features with a ResNet18-style mesh encoder and a ResNet50 image encoder, projects both onto the sphere, and uses reset and forget gates, inspired by GRU, to select reliable features from each source before a mesh decoder produces the depth map. A cache strategy precomputes and stores the face-adjacency information needed for mesh convolutions, avoiding repeated computation. The authors report that SphereFusion achieves the best MRE and MAE on the synthetic 360D dataset and the best MRE and RMSE(log) on Stanford2D3D, with overall accuracy competitive with transformer-based methods, and inference of 0.0174 seconds per 512x1024 panorama.

Load-bearing premise

The speed comparison is the load-bearing premise: SphereFusion's own timing on one RTX 3090 is compared with baseline times reported in other papers, and those baselines may have been measured on different hardware, resolution, or software.

Editorial extensions

If this is right

  • If the reported speed holds, panorama depth estimation becomes practical for real-time robotics and autonomous driving on a single consumer GPU.
  • The accuracy results suggest that a simple ResNet encoder plus a well-chosen projection can rival transformer-based panorama depth models, so future work may not need heavy ViT backbones for this task.
  • The cache strategy makes mesh operations affordable at 512x1024 but, as the paper states, adds GPU memory pressure and currently blocks ultra-high-resolution inputs such as 1024x2048.
  • Because the depth map is produced directly in the spherical domain, SphereFusion avoids the patch-gap artifacts seen in tangent-projection methods like OmniFusion.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication the authors leave implicit is that GateFuse is a generic feature-fusion idea: it could be tested on other 360-degree dense prediction tasks such as semantic segmentation or layout estimation, where the same distortion-versus-detail tradeoff exists.
  • If the fusion is the real source of accuracy, then replacing only the encoders with stronger backbones, while keeping GateFuse and the cache, should improve accuracy without sacrificing the speed advantage.
  • A direct comparison of all methods on identical hardware is the obvious next test; the paper does not report re-running baselines, so the 17 ms figure is best read as a ceiling until such a benchmark exists.
  • The memory cost of the FAF cache suggests the speed gain may not transfer to higher resolutions, so a sparse or on-the-fly adjacency computation would be a natural extension to test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. SphereFusion is a panorama depth estimation method that fuses equirectangular features from a 2D ResNet encoder with spherical-mesh features from a lightweight mesh encoder, using a GRU-inspired GateFuse module to select reliable features in the spherical domain. The paper evaluates the method on 360D, Matterport3D, and Stanford2D3D, reports competitive accuracy against state-of-the-art methods, and claims the fastest inference at 17 ms per 512x1024 panorama on an RTX 3090. Ablation studies isolate the contributions of the two encoders and the fusion strategy.

Significance. If the technical result holds, the paper makes a useful engineering contribution: it shows that a lightweight fusion of equirectangular and spherical features can achieve accuracy close to transformer-based methods while running considerably faster. The evaluation on three public datasets and the ablations isolating encoder and fusion choices are strengths, as is the clear reporting of the authors' own timing procedure. However, the central efficiency claim rests on a timing comparison that is not controlled across baselines, and the GateFuse module is not specified precisely enough for independent implementation. The paper would be a solid contribution after these issues are fixed.

major comments (4)
  1. [Section 4.3, Table 1] The claim that SphereFusion is the most efficient method, requiring only 0.0174 seconds per image during inference, is not supported by a controlled comparison. The text describes the authors' own timing procedure (100 images on a single RTX 3090) but does not state that the baseline timings were re-measured on the same GPU, at the same resolution and precision, or using the same framework. The footnote only mentions recalculating SliceNet metrics from open-source models, not timings. Since Table 1 lists timings from different published papers with possibly different hardware, the headline speed advantage could change under identical conditions. Please re-run all baselines on the same hardware and report the exact setup for each, or, failing that, restrict the claim to 'fastest among published timings rather than a directly measured comparison.'
  2. [Section 4.3 and Section 8] The reported inference time of 0.0174 s does not specify whether it includes the spherical-to-equirectangular (S2E) conversion described in Section 8, which is required to produce the evaluated depth map. Since the network outputs depth in the spherical domain, the end-to-end latency to an equirectangular depth image may be larger than the raw forward pass. Please state clearly whether the 17 ms includes the S2E conversion, the cache construction, and any post-processing needed for evaluation.
  3. [Section 3.3.2] GateFuse, a central contribution, is described only verbally as using a reset gate and a forget gate inspired by GRU. No equations define how the reset gate r and forget gate z are computed from the concatenated features, how they select from Fsp and Feq, or how the final fused feature Ffused is formed. This makes the fusion module underspecified and not independently implementable. Please add the precise gate equations, activation functions, and the fusion formula.
  4. [Section 4.3 vs Section 8] The evaluation depth range is inconsistent between the main text and the supplementary material. Section 4.3 states that pixels outside 0.1-10 meters are ignored for 360D and 0.1-16 meters for Stanford2D3D and Matterport3D, whereas Section 8 of the supplementary states that the depth range is set to 0.1-10 meters. This discrepancy directly affects the computed metrics and comparability with baselines. Please correct the inconsistency and explicitly state the depth range used for each dataset.
minor comments (6)
  1. [Section 4.3] For Matterport3D, the text says the method ranks second on MRE, MAE, and RMSE(log), but Table 1 shows UniFuse has lower values (MRE 0.1063 vs 0.1145, MAE 0.2814 vs 0.2852, RMSE(log) 0.0701 vs 0.0733), so the method actually ranks third on these metrics.
  2. [Supplementary Section 8, Eq. 8] The definitions of MAE and MRE are given as sums without division by N, while RMSE and RMSE(log) divide by N; this is inconsistent and likely a typo. Please insert the missing normalization.
  3. [Section 4.4 and Supplementary Figures] OmniFusion is cited as [51] and PanoFormer as [29] in several places; the correct references are [34] and [42], respectively.
  4. [Throughout] There are several typos and stray fragments: 'SphereFuion' in Section 4.6, 'OminiFusion' in Figure 1, 'Visulization' in Section 9, 'Standard2D3D' in Table 1, and an orphaned citation fragment '[20, 23, 27, 51].' in Section 2.2. Please proofread carefully.
  5. [Section 3.4] The multi-scale loss weights s_i in Eq. 5 are not specified in Section 4.2 or elsewhere; please report the values used during training.
  6. [General] No code is released, which makes it difficult for the community to independently reproduce the accuracy and timing results; releasing code would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: accuracy claims rest on external benchmarks and independent baselines; the uncontrolled timing comparison is a fairness/correctness concern, not a circular derivation.

full rationale

The paper's derivation chain is self-contained with respect to the depth-estimation task. The projections (Eqs. 1, 2, 6), mesh convolution (Eq. 3), and loss functions (Eqs. 4, 5) are standard operations defined from geometry and ground-truth depth, not from the target evaluation metrics. The accuracy comparison uses three public datasets (360D, Matterport3D, Stanford2D3D) and independent published baselines; the ablation studies implement BiFuse and UniFuse fusion inside the same framework and show empirical differences, so GateFuse is not equivalent to its inputs by construction. The only self-citation is to the authors' prior SphereDepth, used as a source of mesh operations and as a baseline; this is not load-bearing because SphereFusion's improvement over SphereDepth is an external empirical result rather than an assumed premise. The 'fastest inference' claim (0.0174 s/image) is based on the authors' own RTX 3090 timing while baseline timings appear to be quoted from original papers without a controlled same-machine re-benchmark; that is a potential benchmarking-fairness weakness, but it is not circular reasoning because no fitted parameter is renamed as a prediction and no equation reduces to its own output. Therefore the paper shows no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

This is a supervised deep learning paper; the central claim is empirical. The ledger captures the hand-set hyperparameters and domain assumptions the method depends on, rather than mathematical axioms or invented physical entities.

free parameters (3)
  • BerHu loss threshold T = 0.2
    Set in Eq. 4 and held constant across all experiments; hand-chosen, not optimized.
  • Spherical mesh resolution ranges = MR 2 to 7 for 360D; MR 3 to 8 for Matterport3D and Stanford2D3D
    Section 3.3.3: the decoder unpooling increases MR from 2 to 7 (360D) or 3 to 8 (other datasets); these ranges determine triangle counts and are chosen to match input resolution.
  • Multi-scale loss weights s_i = unspecified
    Eq. 5 introduces per-scale weights s_i but the paper never gives their values; this is an unstated hyperparameter that affects training.
assumptions (3)
  • domain assumption The icosahedral spherical mesh at the chosen MRs is an adequate approximation of the sphere for depth estimation without significant distortion or discontinuity artifacts.
    Section 3.1: the paper states the spherical mesh approximates the sphere and that the difference becomes smaller with higher MR, but provides no quantitative guarantee that MR 7/8 is sufficient.
  • domain assumption E2S bilinear sampling and S2E nearest-triangle assignment preserve the correspondence between equirectangular and spherical features well enough for fusion.
    Supplementary Section 7 defines E2S and S2E; the fusion module implicitly assumes the projected features from the two domains refer to the same 3D surface locations.
  • domain assumption The evaluation protocol, including depth range clipping (0.1 to 10 m for 360D, 0.1 to 16 m for the real datasets), matches the protocols used in the cited baseline papers so that the published numbers are comparable.
    Section 4.3 describes the depth range filtering following BiFuse, but it is not verified that every baseline paper used identical filtering and resolution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SphereFusion: Efficient Panorama Depth Estimation via Gated Fusion." pith.science (2026). https://pith.science/paper/WN66TYLJ

@misc{pith2026250205859,
  author       = {Pith},
  title        = {Pith review of: SphereFusion: Efficient Panorama Depth Estimation via Gated Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WN66TYLJ}},
  note         = {Machine review of arXiv:2502.05859}
}
abstract

Due to the rapid development of panorama cameras, the task of estimating panorama depth has attracted significant attention from the computer vision community, especially in applications such as robot sensing and autonomous driving. However, existing methods relying on different projection formats often encounter challenges, either struggling with distortion and discontinuity in the case of equirectangular, cubemap, and tangent projections, or experiencing a loss of texture details with the spherical projection. To tackle these concerns, we present SphereFusion, an end-to-end framework that combines the strengths of various projection methods. Specifically, SphereFusion initially employs 2D image convolution and mesh operations to extract two distinct types of features from the panorama image in both equirectangular and spherical projection domains. These features are then projected onto the spherical domain, where a gate fusion module selects the most reliable features for fusion. Finally, SphereFusion estimates panorama depth within the spherical domain. Meanwhile, SphereFusion employs a cache strategy to improve the efficiency of mesh operation. Extensive experiments on three public panorama datasets demonstrate that SphereFusion achieves competitive results with other state-of-the-art methods, while presenting the fastest inference speed at only 17 ms on a 512$\times$1024 panorama image.

Figures

Figures reproduced from arXiv: 2502.05859 by the authors.

Figure 1
Figure 1. Comparison with BiFuse [48], UniFuse [29], SliceNet [37], PanoFormer [42], OminiFusion [34], SphereDepth [51], Ho￾hoNet [46] on Stanford2D3D [3] with resolution of 512 × 1024. The horizontal axis is the FPS, and the vertical axis is δ(1.25)(%), which counts the percentage of the absolute relative difference be￾tween the prediction and the ground truth that is less than 1.25. The higher FPS and higher δ(1.25)(%) mean… view at source ↗
Figure 2
Figure 2. Given a panorama image in the equirectangular projection and the spherical projection, SphereFusion simultaneously extracts [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The ideal representation of a panorama image is the sphere, but it is impractical. The equirectangular projection is the most [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Mesh Operations includes Mesh Convolution and the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: We implement BiFuse [48], UniFuse [29], and our GateFuse to fuse features from spherical projection Fsp and equirectangular projection Feq. Unlike BiFuse and UniFuse select features from Feq and fuse them to Fsp, GateFuse selects features from Fsp and Feq. 3.3.2 The Fu…
Figure 6
Figure 6. Figure 6: We select two scenes from 360D [59] and visualize depth maps and point clouds. RGB GT SliceNet PanoFormer ours GT SliceNet PanoFormer ours [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: We select two scenes from Matterport3D [ [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: We select two scenes from Stanford2D3D [ [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: We select two scenes from 360D [59] and compare depth maps from different fusion strategies. Our GateFuse can recon￾struct more details, and we mark out these regions with red boxes. 4.6. Limitations We propose SphereFuion for panorama depth estimation by using the 2D …
Figure 10
Figure 10. Figure 10: Depth Maps of 360D. Invalid parts of the depth map are set to red. SphereDepth [51] UniFuse [29] SphereFusion (ours) [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Point Clouds of 360D. Our method reconstructs more details of the scene [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Depth Maps of Matterport3D. Invalid parts of the depth map are set to red. OmniFusion [34] PanoFormer [42] SphereFusion (ours) [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Depth Maps of Matterport3D. Our method has less noise and maintains the structure of the scene [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Depth Maps of Stanford2D3D. Invalid parts of the depth map are set to red. OmniFusion [34] PanoFormer [42] SphereFusion (ours) [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Depth Maps of Stanford2D3D. Our method does not suffer from discontinuity [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 48 canonical work pages

  1. [1]

    Pano3d: A holistic benchmark and a solid baseline for 360deg depth estimation

    Georgios Albanis, Nikolaos Zioulis, Petros Drakoulis, Vasileios Gkitsas, Vladimiros Sterzentsenko, Federico Al- varez, Dimitrios Zarpalas, and Petros Daras. Pano3d: A holistic benchmark and a solid baseline for 360deg depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3727– 3737, 2021. 1

    work page 2021

  2. [2]

    High quality monoc- ular depth estimation via transfer learning

    Ibraheem Alhashim and Peter Wonka. High quality monoc- ular depth estimation via transfer learning. arXiv preprint arXiv:1812.11941, 2018. 2

    arXiv 2018

  3. [3]

    Joint 2d-3d-semantic data for indoor scene understanding

    Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 1, 6, 7

    arXiv 2017

  4. [4]

    Adabins: Depth estimation using adaptive bins

    Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009–4018, 2021. 2

    work page 2021

  5. [5]

    Unsupervised scale-consistent depth learning from video.In- ternational Journal of Computer Vision, 129(9):2548–2564,

    Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Zhichao Li, Le Zhang, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsupervised scale-consistent depth learning from video.In- ternational Journal of Computer Vision, 129(9):2548–2564,

    work page

  6. [6]

    Matterport3d: Learning from rgb-d data in indoor environments

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 1, 6, 7

    arXiv 2017

  7. [7]

    Distortion-aware monocular depth estimation for omnidirectional images

    Hong-Xiang Chen, Kunhong Li, Zhiheng Fu, Mengyi Liu, Zonghao Chen, and Yulan Guo. Distortion-aware monocular depth estimation for omnidirectional images. IEEE Signal Processing Letters, 28:334–338, 2021. 2, 3

    work page 2021

  8. [8]

    Cube padding for weakly-supervised saliency prediction in 360 videos

    Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao- Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in 360 videos. InPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1420–1429, 2018. 2

    work page 2018

Show all 63 references
  1. [9]

    Omnidirectional depth extension networks

    Xinjing Cheng, Peng Wang, Yanqi Zhou, Chenye Guan, and Ruigang Yang. Omnidirectional depth extension networks. In 2020 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 589–595. IEEE, 2020. 3

    2020

  2. [10]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation

    Kyunghyun Cho, Bart Van Merri ¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 5

    2014 arXiv

  3. [11]

    MeshLab: an Open-Source Mesh Processing Tool

    Paolo Cignoni, Marco Callieri, Massimiliano Corsini, Mat- teo Dellepiane, Fabio Ganovelli, and Guido Ranzuglia. MeshLab: an Open-Source Mesh Processing Tool. In Euro- graphics Italian Chapter Conference. The Eurographics As- sociation, 2008. 7

    2008

  4. [12]

    Spherical cnns

    Taco S Cohen, Mario Geiger, Jonas K ¨ohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130,

    arXiv

  5. [13]

    Eliminating the blind spot: Adapting 3d object detection and monocular depth estimation to 360 panoramic imagery

    Greire Payen de La Garanderie, Amir Atapour Abarghouei, and Toby P Breckon. Eliminating the blind spot: Adapting 3d object detection and monocular depth estimation to 360 panoramic imagery. In Proceedings of the European Con- ference on Computer Vision (ECCV), pages 789–807, 2018. 3

    2018

  6. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  7. [15]

    Pano popups: In- door 3d reconstruction with a plane-aware network

    Marc Eder, Pierre Moulon, and Li Guan. Pano popups: In- door 3d reconstruction with a plane-aware network. In 2019 International Conference on 3D Vision (3DV), pages 76–84. IEEE, 2019. 3

    2019

  8. [16]

    Mapped convolutions

    Marc Eder, True Price, Thanh Vu, Akash Bapat, and Jan- Michael Frahm. Mapped convolutions. arXiv preprint arXiv:1906.11096, 2019. 3

    1906 arXiv

  9. [17]

    Tangent images for mitigating spherical distortion

    Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12426–12434, 2020. 1, 3, 4

    2020

  10. [18]

    Predicting depth, surface nor- mals and semantic labels with a common multi-scale con- volutional architecture

    David Eigen and Rob Fergus. Predicting depth, surface nor- mals and semantic labels with a common multi-scale con- volutional architecture. In Proceedings of the IEEE inter- national conference on computer vision , pages 2650–2658,

  11. [19]

    Depth map prediction from a single image using a multi-scale deep net- work

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep net- work. Advances in neural information processing systems , 27, 2014. 2

    2014

  12. [20]

    Meshnet: Mesh neural network for 3d shape rep- resentation

    Yutong Feng, Yifan Feng, Haoxuan You, Xibin Zhao, and Yue Gao. Meshnet: Mesh neural network for 3d shape rep- resentation. In Proceedings of the AAAI Conference on Arti- ficial Intelligence, pages 8279–8286, 2019. 2, 3

    2019

  13. [21]

    Corners for layout: End-to-end layout recovery from 360 images

    Clara Fernandez-Labrador, Jose M Facil, Alejandro Perez- Yus, C´edric Demonceaux, Javier Civera, and Jose J Guer- rero. Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters , 5(2): 1255–1262, 2020. 3

    2020

  14. [22]

    Digging into self-supervised monocular depth estimation

    Cl ´ement Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 3828–3838,

  15. [23]

    Meshcnn: a network with an edge

    Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshcnn: a network with an edge. ACM Transactions on Graphics (TOG), 38(4):1–12,

  16. [24]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 2, 3, 4

    2016

  17. [25]

    Auto- matic photo pop-up

    Derek Hoiem, Alexei A Efros, and Martial Hebert. Auto- matic photo pop-up. ACM Trans. Graph., pages 577–584,

  18. [26]

    Jittor: a novel deep learning framework with meta-operators and unified graph execution

    Shi-Min Hu, Dun Liang, Guo-Ye Yang, Guo-Wei Yang, and Wen-Yang Zhou. Jittor: a novel deep learning framework with meta-operators and unified graph execution. Science China Information Sciences, 63(222103):1–21, 2020. 6

    2020

  19. [27]

    Subdivision-based mesh convolution networks

    Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Trans- actions on Graphics (TOG), 41:1–16, 2022. 2, 3, 4

    2022

  20. [28]

    Spherical cnns on unstructured grids

    Chiyu Jiang, Jingwei Huang, Karthik Kashinath, Philip Mar- cus, Matthias Niessner, et al. Spherical cnns on unstructured grids. arXiv preprint arXiv:1901.02039, 2019. 3

    1901 arXiv

  21. [29]

    Unifuse: Unidirectional fusion for 360 panorama depth estimation

    Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519–1526, 2021. 1, 2, 3, 5, 6, 7, 8

    2021

  22. [30]

    Geometric structure based and regularized depth estimation from 360 indoor imagery

    Lei Jin, Yanyu Xu, Jia Zheng, Junfei Zhang, Rui Tang, Shugong Xu, Jingyi Yu, and Shenghua Gao. Geometric structure based and regularized depth estimation from 360 indoor imagery. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 889...

    2020

  23. [31]

    Himode: A hybrid monocular omnidirectional depth estimation model

    Masum Shah Junayed, Arezoo Sadeghzadeh, Md Baharul Is- lam, Lai-Kuan Wong, and Tarkan Aydın. Himode: A hybrid monocular omnidirectional depth estimation model. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5212–5221, 2022. 2

    2022

  24. [32]

    Deeper depth prediction with fully convolutional residual networks

    Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Fed- erico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV) , pages 239–

    2016

  25. [33]

    Spherephd: Applying cnns on 360° images with non-euclidean spherical polyhedron represen- tation

    Yeonkun Lee, Jaeseok Jeong, Jongseob Yun, Wonjune Cho, and Kuk-Jin Yoon. Spherephd: Applying cnns on 360° images with non-euclidean spherical polyhedron represen- tation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3

    2020

  26. [34]

    Omnifusion: 360 monocular depth estimation via geometry-aware fusion

    Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2801–2810, 2022. 1, 2, 3, 6, 7, 4

    2022

  27. [35]

    P3depth: Monocular depth estimation with a piecewise planarity prior

    Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1610–1621, 2022. 2

    2022

  28. [36]

    High-resolution depth estimation for 360-degree panoramas through perspective and panoramic depth images registration

    Chi-Han Peng and Jiayao Zhang. High-resolution depth estimation for 360-degree panoramas through perspective and panoramic depth images registration. arXiv preprint arXiv:2210.10414, 2022. 2, 3

    2022 arXiv

  29. [37]

    Slicenet: deep dense depth esti- mation from a single indoor panorama using a slice-based representation

    Giovanni Pintore, Marco Agus, Eva Almansa, Jens Schnei- der, and Enrico Gobbetti. Slicenet: deep dense depth esti- mation from a single indoor panorama using a slice-based representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

    2021

  30. [38]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 2

    2020

  31. [39]

    360monodepth: High-resolution 360deg monocular depth estimation

    Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3762– 3772, 2022. 3, 8

    2022

  32. [40]

    U- net: Convolutional networks for biomedical image segmen- tation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- net: Convolutional networks for biomedical image segmen- tation. In International Conference on Medical image com- puting and computer-assisted intervention , pages 234–241. Springer, 2015. 5

    2015

  33. [41]

    Make3d: Learning 3d scene structure from a single still image

    Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image. IEEE transactions on pattern analysis and machine intelligence , 31(5):824–840, 2008. 2

    2008

  34. [42]

    Panoformer: Panorama transformer for indoor 360 depth estimation

    Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panoformer: Panorama transformer for indoor 360 depth estimation. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part I, pages 195–211. Springer,

    2022

  35. [43]

    Neural contourlet network for monocular 360 depth estimation

    Zhijie Shen, Chunyu Lin, Lang Nie, Kang Liao, and Yao Zhao. Neural contourlet network for monocular 360 depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8574–8585, 2022. 2

    2022

  36. [44]

    Standardization status of 360 degree video coding and delivery

    Robert Skupin, Yago Sanchez, Y-K Wang, Miska M Han- nuksela, J Boyce, and Mathias Wien. Standardization status of 360 degree video coding and delivery. In 2017 IEEE Vi- sual Communications and Image Processing (VCIP) , pages 1–4. IEEE, 2017. 1

    2017

  37. [45]

    Kernel transformer net- works for compact spherical convolution

    Yu-Chuan Su and Kristen Grauman. Kernel transformer net- works for compact spherical convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442–9451, 2019. 3

    2019

  38. [46]

    Hohonet: 360 indoor holistic understanding with latent horizontal fea- tures

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal fea- tures. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2573–2582,

  39. [47]

    Distortion-aware convolutional filters for dense prediction in panoramic images

    Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In Proceedings of the European Confer- ence on Computer Vision (ECCV), pages 707–722, 2018. 2, 3

    2018

  40. [48]

    Bifuse: Monocular 360 depth estimation via bi-projection fusion

    Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 1, 2, 3, 5, 6, 7, 8

    2020

  41. [49]

    360sd-net: 360 stereo depth estimation with learnable cost volume

    Ning-Hsu Wang, Bolivar Solarte, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. 360sd-net: 360 stereo depth estimation with learnable cost volume. In2020 IEEE International Con- ference on Robotics and Automation (ICRA), pages 582–588. IEEE, 2020. 3

    2020

  42. [50]

    The temporal opportunist: Self-supervised multi-frame monocular depth

    Jamie Watson, Oisin Mac Aodha, Victor Prisacariu, Gabriel Brostow, and Michael Firman. The temporal opportunist: Self-supervised multi-frame monocular depth. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1164–1174, 2021. 2

    2021

  43. [51]

    Spheredepth: Panorama depth estima- tion from spherical domain

    Qingsong Yan, Qiang Wang, Kaiyong Zhao, Bo Li, Xiaowen Chu, and Fei Deng. Spheredepth: Panorama depth estima- tion from spherical domain. In2022 Tenth international con- ference on 3D vision (3DV), 2022. 1, 2, 3, 4, 6, 7, 8

    2022

  44. [52]

    Virtual normal: En- forcing geometric constraints for accurate and robust depth prediction

    Wei Yin, Yifan Liu, and Chunhua Shen. Virtual normal: En- forcing geometric constraints for accurate and robust depth prediction. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2021. 2

    2021

  45. [53]

    Flat2sphere: Learning spheri- cal convolution for fast features from 360 imagery

    S Yu-Chuan and G Kristen. Flat2sphere: Learning spheri- cal convolution for fast features from 360 imagery. In Pro- ceedings of International Conference on Neural Information Processing Systems (NIPS), 2017. 3

    2017

  46. [54]

    New crfs: Neural window fully-connected crfs for monocular depth estimation

    Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. New crfs: Neural window fully-connected crfs for monocular depth estimation. arXiv preprint arXiv:2203.01502, 2022. 2

    2022 arXiv

  47. [55]

    Joint 3d layout and depth prediction from a single indoor panorama image

    Wei Zeng, Sezer Karaoglu, and Theo Gevers. Joint 3d layout and depth prediction from a single indoor panorama image. In European Conference on Computer Vision , pages 666–

  48. [56]

    To- wards better generalization: Joint depth-pose learning with- out posenet

    Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. To- wards better generalization: Joint depth-pose learning with- out posenet. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9151– 9161, 2020. 2

    2020

  49. [57]

    Structured3d: A large photo-realistic dataset for structured 3d modeling

    Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In European Conference on Computer Vision, pages 519–535. Springer, 2020. 1

    2020

  50. [58]

    Unsupervised learning of depth and ego-motion from video

    Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851–1858, 2017. 2

    2017

  51. [59]

    Omnidepth: Dense depth estimation for indoors spherical panoramas

    Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omnidepth: Dense depth estimation for indoors spherical panoramas. In Proceedings of the Euro- pean Conference on Computer Vision (ECCV) , pages 448– 465, 2018. 1, 2, 3, 4, 6, 7, 8 SphereFusion: Effici...

    2018

  52. [60]

    Panorama Projections To capture the texture features and avoid distortion and discontinuity, SphereFusion relies on two panorama pro- jections: the equirectangular projection and the spherical mesh. In this section, we describe the details of their con- version, and how to con...

  53. [61]

    Fol- lowing BiFuse [48], we use five evaluation metrics, includ- ing MAE, MRE, RMSE, RMSE(log), and δn

    Evaluation To compare with other methods, we convert our depth map in the spherical domain to equirectangular projection. Fol- lowing BiFuse [48], we use five evaluation metrics, includ- ing MAE, MRE, RMSE, RMSE(log), and δn. Eq. 8 shows how to calculate them, where gt is the ...

  54. [62]

    On 360D [59], we compare our method with SphereDepth [51] and UniFuse [29]

    Visulization In this section, we add more visualization results on three datasets and compare SphereFusion (ours) with state-of- the-art methods. On 360D [59], we compare our method with SphereDepth [51] and UniFuse [29]. Figure 10 and Figure 11 show depth maps and point cloud...

  55. [63]

    Figure 12 and Fig- ure 13 show depth maps generated by different methods on Matterport3D and corresponding point clouds

    and Stanford2D3D [3], we compare our method with OmniFusion [51] and PanoFormer [29]. Figure 12 and Fig- ure 13 show depth maps generated by different methods on Matterport3D and corresponding point clouds. Figure 14 and Figure 15 show results generated by different meth- ods ...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.