Pith. sign in

REVIEW 4 major objections 6 minor 45 references

M3D: Dual-Stream Selective State Spaces and Depth-Driven Framework for High-Fidelity Single-View 3D Reconstruction

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A dual-stream architecture that combines a selective state-space model with a depth branch is claimed to achieve state-of-the-art single-view 3D reconstruction, with a 36.9% lower Chamfer Distance and 13.3% higher F-score than the SSR…

desk verdict Plausible architecture, unsupported SOTA claim: Table III lacks protocol and inner numbers conflict. read the letter →

arxiv 2411.12635 v2 pith:OF3NLZ2F submitted 2024-11-19 cs.CV

Luoxi Zhang , Pragyan Shrestha , Yu Zhou , Chun Xie , Itaru Kitahara This is my paper · ORCID
classification cs.CV
keywords single-view3Dreconstructionimplicitneuralrepresentationselectivestatespacemodeldepthestimationdual-streamfeatureextraction3D-FRONTdatasetsigneddistancefunctionvision
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

This paper tries to establish that a single-view 3D reconstruction network can achieve high-fidelity indoor object geometry by processing RGB and depth in two separate streams and fusing them with a selective state-space model. On the 3D-FRONT dataset, the proposed M3D framework reports a Chamfer Distance of 6.60, an F-score of 80.85, and a Normal Consistency of 0.901, corresponding to a 36.9% CD improvement, 13.3% F-score gain, and 5.5% NC gain over the SSR baseline. The motivation is that CNNs capture local details but miss global context, while transformers capture global context at quadratic cost; the paper argues a selective SSM plus explicit depth cues gets both. If correct, the design gives a practical recipe for preserving fine details and occlusion robustness in indoor scene reconstruction for VR, robotics, and driving.

What carries the argument

M3D's load-bearing mechanism is the dual-stream Selective Attention Module combined with a depth-driven geometric stream. In the RGB stream, high-dimensional features are channel-split: one half is processed by a selective state-space model (a sequence model that scans tokens in linear time and selectively retains relevant context), the other by convolutional residual blocks, and the two are merged and refined by self-attention; this iteration is repeated with fewer layers to refine global and local features. In parallel, a pretrained monocular depth estimator produces a depth map offline, and the depth feature is combined with shallow RGB features via generalized addition and bilinear interpolation onto 2D-projected 3D points, then concatenated with the point coordinate and passed through an MLP. The implicit SDF is converted to density via a learnable $\beta$ for differentiable volume rendering, and the loss combines 3D SDF, color, depth consistency, and normal consistency terms.

What would settle it

Run M3D and the listed comparison methods on one shared 3D-FRONT test image using identical mesh extraction and metric code, and check whether M3D's Chamfer Distance is around 6.60 (as in the main table) or around 0.0066 (as in the large-model comparison); reproducing either value, and seeing whether the baselines reproduce their published values, settles which comparison is meaningful.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the global/local feature tradeoff in single-view 3D reconstruction is best resolved by a dual-stream design: one stream runs a Selective Attention Module that splits features and processes half with a selective state-space model and half with convolutional residual blocks before self-attention, while a parallel stream injects an offline-estimated depth map through generalized addition and bilinear projection onto 3D points. The fused RGB-depth feature is decoded by an implicit signed-distance-field representation with volumetric rendering and supervised by geometry, color, depth, and normal losses. The paper reports that the depth branch alone reduces Chamfer Distance by 28.6% relative to the SSR baseline at epoch 120, and the full M3D reduces it by 53.9%, with the selective SSM credited for smoother normals and faster convergence. The claimed result is state-of-the-art reconstruction fidelity on the 3D-FRONT benchmark across CD, F-score, NC, PSNR, and IoU.

Load-bearing premise

The central claim assumes the comparison numbers for other methods in Table III were produced on the same evaluation set, resolution, and metric implementation as M3D's numbers; if they were not, the reported state-of-the-art advantage is not established.

Editorial extensions

If this is right

  • If the reported gains hold, separating RGB and depth streams is a more effective way to use monocular depth in neural implicit reconstruction than joint single-stream feature extraction.
  • The linear-time selective SSM component suggests the framework can scale to higher image resolutions than transformer-only backbones without quadratic attention cost.
  • Offline depth estimation keeps training costs low while still providing geometric cues, so the architecture is feasible on a single GPU with batch size 30.
  • Each added component (enhanced residual blocks, depth branch, selective attention) contributes positively in ablations, implying the gains are compositional rather than from one module alone.

Reading between the lines

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

  • Because the depth stream leans on a specific pretrained estimator, M3D's fidelity gains may track the accuracy of that estimator; swapping in other monocular depth models would test how much of the improvement is depth quality versus architecture.
  • The same dual-stream fusion could transfer to multi-view or video reconstruction, where depth is easier to obtain and the SSM's linear scan could aggregate temporal context.
  • The 3D-FRONT results are for indoor furniture categories; applying M3D to outdoor or deformable objects would clarify whether the benefit is specific to indoor geometry.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. The manuscript proposes M3D, a single-view 3D reconstruction method that uses a dual-stream architecture: an RGB stream built on selective state-space models with residual and attention components, and a depth stream based on pre-trained monocular depth maps. The two streams are fused with an implicit SDF representation and volumetric rendering, trained with a two-stage loss. The paper claims state-of-the-art performance on 3D-FRONT, with headline gains of 36.9% in Chamfer Distance, 13.3% in F-Score, and 5.5% in Normal Consistency over SSR, and also reports favorable comparisons with recent large reconstruction models. The contribution is primarily architectural: balancing global and local feature extraction through a dual-stream SSM design while injecting depth information.

Significance. The architectural idea is credible and the ablation sequence directly targets the stated design choices: enhanced shallow residual blocks, a depth branch, and an SSM-based deep feature module. Training hyperparameters and loss weights are disclosed, which aids reproducibility. However, the central SOTA claim is not supported by the current evidence: the comparison with recent models in Table III is not a controlled experiment, and several internal inconsistencies in the reported tables make the quantitative claims unreliable. The qualitative figures cannot compensate for these problems. If the authors can supply a common evaluation protocol and reconcile the tables, the method would be a plausible contribution, but as written the empirical case is not established.

major comments (4)
  1. [Section IV-B, Table III] The comparison in Table III is not controlled and cannot support the central SOTA claim. The table does not state the dataset, input resolution, rendering protocol, mesh extraction settings, CD normalization, or whether the baselines were retrained or evaluated under the same conditions as M3D. The same model reports CD 6.60 on 3D-FRONT in Table I but CD 0.0066 in Table III, and the baseline PSNR values all lie about 10 dB below M3D's 30.04. These signatures strongly suggest that the numbers were not produced on the same benchmark. Since Table III is the only comparison against OpenLRM, InstantMesh, CRM, Unique3D, and Wonder3D, the state-of-the-art conclusion is unsupported unless all baselines are rerun under one common protocol and that protocol is specified in the paper.
  2. [Section V-B, Table V; Section I; Table I] The ablation results in Table V contradict the headline improvements. Table V reports M3D's CD as 7.61 at epoch 120 and the SSR baseline as 16.51, which is a 53.9% improvement, while Section I and Table I report CD 6.60 versus 10.45, i.e., 36.9%. The same mismatch appears for F-Score (32.7% vs 13.3%) and NC (12.7% vs 5.5%). If the two baselines are different training checkpoints, evaluation conditions, or splits, that must be stated; otherwise the paper's reported effect sizes are internally inconsistent and the ablation conclusions are not reliable.
  3. [Table I] The 'mean' column in Table I is not reproducible from the eight per-category numbers shown. For example, the arithmetic mean of the SSR CD row is 11.32, not the reported 10.45; for M3D it is 8.88, not 6.60; several F-Score and NC rows show similar mismatches. If the means are weighted by object frequency or by some other factor, the weighting must be given; without this, the percentage improvements quoted in the abstract and Section I cannot be verified.
  4. [Section IV-B, Table II] Table II compares M3D with Zero-1-to-3 and Shape-E without stating the dataset or evaluation protocol. The reported M3D values (CD 6.60, F-Score 80.85, NC 0.901) match Table I's 3D-FRONT numbers, but no evidence is provided that Zero-1-to-3 and Shape-E were evaluated on the same 3D-FRONT split with the same metrics. This comparison is load-bearing because it is used to argue superiority over methods with large-scale pre-trained priors.
minor comments (6)
  1. [Abstract and Section I] The paper states that code and dataset are available at 'this URL', but no URL appears anywhere in the manuscript; please provide the actual repository link.
  2. [Section IV-B] There is a sentence fragment beginning 'blocks [7, 8] for shallow feature extraction contribute to a 56% reduction in CD'; the preceding text appears to have been cut and should be rewritten.
  3. [Equations (6) and (13)] The symbol beta is used both as the learnable SDF-to-density temperature in Eq. (6) and as a loss weight in Eq. (13); please use distinct symbols.
  4. [Figure 4 caption] The caption states that F-Score values are divided by 100 for visualization, but the F-Score values in Table V are already in [0,1]; please clarify the scaling so the axes can be interpreted.
  5. [Table I] The column headers 'CD ↓ (51.0%)', 'F-Score ↑ (33.6%)', and 'NC ↑ (10.3%)' are unexplained; if these are aggregation weights or other factors, define them in the caption.
  6. [References] References [2] and [18] share the same title and are cited in similar contexts; please verify that they are distinct works and update the citation labels and reference list accordingly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported SOTA comparisons are uncontrolled but the derivation chain is not circular.

full rationale

This is an empirical systems paper with no formal derivation chain whose conclusions are loaded onto a self-citation or a fitted parameter. The core experiments train M3D on the 3D-FRONT split of Liu et al. and compare, under the same protocol in Tables I and V, against a reproduced SSR baseline and against published numbers for MGN, LIEN, InstPIFu, Zero-1-to-3, and Shape-E. The ablations in Table V compare model variants trained in the same setup, so the incremental CD/F/NC gains are not figure-ground illusions: each row is a distinct configuration, not a renamed version of the same fitted output. The depth branch uses an offline pre-trained depth estimator as input and as auxiliary supervision; this is a design choice, not a self-definitional prediction, because the reconstructed SDF and mesh are still supervised by ground-truth geometry and RGB, not defined as the depth prior. No load-bearing argument cites the authors' own prior work; references [3, 7, 8, 9, 41] are external. There are real comparability and consistency problems outside circularity: Table III states no dataset, resolution, or metric protocol, and its CD value (0.0066) and PSNR gap (~10 dB) are implausible relative to Table I (CD 6.60); the paper also reports a 36.9% CD improvement in the introduction but 53.9% in Table V. Those are correctness and benchmark-integrity concerns, not circular reasoning, and under the review rules they do not raise the circularity score. I therefore find no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

No new entities are postulated. The central claim rests on standard deep-learning components and several unstated assumptions about evaluation fairness and depth-map quality. Free parameters are hand-set hyperparameters; the only learned extra parameter is the SDF-to-density temperature. The most consequential unstated assumptions are that Depth Anything's relative depth maps can be treated as precise geometric cues without scale or shift alignment, and that published numbers for baselines in Tables II and III were produced under the same protocol.

free parameters (4)
  • loss weights alpha, beta, gamma, delta = alpha=1, beta=0.1, gamma=0.1, delta=0.01
    Set by hand in Section III-D; no sensitivity analysis is reported.
  • learning rate schedule = initial 0.00006, dropped to 0.00001 at epoch 110
    Chosen for SSM training in Section IV-A; no search details are given.
  • selective attention module layer counts = N1=8 layers, N2=4 layers
    Set in Section III-B; no ablation varies these layer counts.
  • SDF-to-density temperature beta = learned
    Learnable parameter in Eq. 6, trained with the network; not an independent fit, but it is an extra model parameter.
assumptions (4)
  • domain assumption Volumetric rendering equations (Eq. 6-8) correctly approximate light transport and are differentiable.
    Adopted from NeRF/NeuS-style literature; M3D does not re-derive or validate them for this setting.
  • domain assumption Depth maps from the pretrained Depth Anything model are accurate and aligned with the 3D-FRONT coordinate frame.
    Used in Section III-B to create the geometric stream; the paper reports no validation of depth quality, scale, or shift alignment. Depth Anything outputs affine-invariant rather than metric depth.
  • domain assumption The 3D-FRONT train/test split and ground-truth SDF supervision are correct and identical to those used for the baselines.
    Section IV-A says it adopts Liu et al.'s splits; no verification is provided for the comparison methods.
  • ad hoc to paper Published baseline numbers in Tables II and III come from the same evaluation protocol as M3D.
    No protocol is stated, and the large PSNR gap suggests otherwise; this assumption is necessary for the SOTA claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of M3D: Dual-Stream Selective State Spaces and Depth-Driven Framework for High-Fidelity Single-View 3D Reconstruction." pith.science (2026). https://pith.science/paper/OF3NLZ2F

@misc{pith2026241112635,
  author       = {Pith},
  title        = {Pith review of: M3D: Dual-Stream Selective State Spaces and Depth-Driven Framework for High-Fidelity Single-View 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OF3NLZ2F}},
  note         = {Machine review of arXiv:2411.12635}
}
read the original abstract

The precise reconstruction of 3D objects from a single RGB image in complex scenes presents a critical challenge in virtual reality, autonomous driving, and robotics. Existing neural implicit 3D representation methods face significant difficulties in balancing the extraction of global and local features, particularly in diverse and complex environments, leading to insufficient reconstruction precision and quality. We propose M3D, a novel single-view 3D reconstruction framework, to tackle these challenges. This framework adopts a dual-stream feature extraction strategy based on Selective State Spaces to effectively balance the extraction of global and local features, thereby improving scene comprehension and representation precision. Additionally, a parallel branch extracts depth information, effectively integrating visual and geometric features to enhance reconstruction quality and preserve intricate details. Experimental results indicate that the fusion of multi-scale features with depth information via the dual-branch feature extraction significantly boosts geometric consistency and fidelity, achieving state-of-the-art reconstruction performance.

Figures

Figures reproduced from arXiv: 2411.12635 by the authors.

Figure 1
Figure 1. Overview of the M3D framework, showing the dual-stream architecture, selective attention module, implicit representation, and rendering network. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of 3D reconstruction results from different methods. Columns from left to right represent the input images, results from MGN[16], [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Texture comparison between M3D and SSR [35], highlighting M3D’s superior texture smoothness and color accuracy. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation experiments comparing the performance of different [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 33 canonical work pages

  1. [1]

    and Author D

    Author C. and Author D. Graph cut optimization for 3d shape reconstruction. In Proceedings of the International Conference on Computer Vision (ICCV) , pages 150–160, 2017

    work page 2017

  2. [2]

    Y . Chen, M. Fang, and R. Zhao. Towards high-fidelity single- view holistic reconstruction of indoor scenes. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 340–351, 2020

    work page 2020

  3. [3]

    Transformers are ssms: Generalized models and efficient algorithms through structured state space duality, 2024

    Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality, 2024

    2024

  4. [4]

    Geogen: Geometry-aware generative modeling via signed distance functions, 2024

    Salvatore Esposito, Qingshan Xu, Kacper Kania, Charlie He- witt, Octave Mariotti, Lohit Petikam, Julien Valentin, Arno Onken, and Oisin Mac Aodha. Geogen: Geometry-aware generative modeling via signed distance functions, 2024

    work page 2024

  5. [5]

    A point set generation network for 3d object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 605–613, 2017

    work page 2017

  6. [6]

    3d-front: 3d furnished rooms with layouts and semantics

    Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In International Conference on Computer Vision (ICCV), 2021

    work page 2021

  7. [7]

    Mambavision: A hy- brid mamba-transformer vision backbone

    Ali Hatamizadeh and Jan Kautz. Mambavision: A hy- brid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024

    arXiv 2024

  8. [8]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778, 2016

    work page 2016

Show all 45 references
  1. [9]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778, 2016

    2016

  2. [10]

    Openlrm: Open-source large reconstruction models, 2023

    Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models, 2023

    2023

  3. [11]

    Shap-e: Generating conditional 3d implicit functions

    Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463 , 2023

    2023 arXiv

  4. [12]

    Repurposing diffusion-based image generators for monocular depth estima- tion, 2024

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estima- tion, 2024

    2024

  5. [13]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (TOG) , 36(4): 1–13, 2017

    2017

  6. [14]

    Im- agenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Im- agenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NIPS) , pages 1097–1105, 2012

    2012

  7. [15]

    Recent advances in multi-modal 3d scene understanding: A comprehensive survey and evaluation, 2023

    Yinjie Lei, Zixuan Wang, Feng Chen, Guoqing Wang, Peng Wang, and Yang Yang. Recent advances in multi-modal 3d scene understanding: A comprehensive survey and evaluation, 2023

    2023

  8. [16]

    J. Li, X. Wang, and D. Li. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 55– 65, 2020

    2020

  9. [17]

    Planercnn: 3d plane detection and reconstruction from a single image

    Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4445– 4454, 2019

    2019

  10. [18]

    Towards high-fidelity single-view holistic reconstruction of indoor scenes

    Haolin Liu, Yujian Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision (ECCV) , 2022

    2022

  11. [19]

    Zero-1-to-3: Zero-shot one image to 3d object, 2023

    Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl V ondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023

    2023

  12. [20]

    Theobalt, Christian

    Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, and et al. Theobalt, Christian. Wonder3d: Single image to 3d using cross-domain diffusion. arXiv preprint arXiv:2310.15008, 2023

    arXiv

  13. [21]

    Marching cubes: A high resolution 3d surface construction algorithm

    William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM Transactions on Graphics (TOG) , 21(4):163–169, 1987

    1987

  14. [22]

    Occupancy networks: Learning 3d reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebas- tian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Conference on Computer Vision and Pattern Recognition (CVPR) , 2019

    2019

  15. [23]

    Occupancy networks: Learning 3d reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebas- tian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 4460–4470, 2019

    2019

  16. [24]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthe- sis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405–421, 2020

    2020

  17. [25]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard New- combe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 165–174, 2019

    2019

  18. [26]

    V olumetric and multi-view cnns for object classification on 3d data

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. V olumetric and multi-view cnns for object classification on 3d data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 5648–5656, 2016

    2016

  19. [27]

    Midas: High-quality depth estimation with minimal training data

    Rene Ranftl and et al. Midas: High-quality depth estimation with minimal training data. In Proceedings of the European Conference on Computer Vision (ECCV) , 2020

    2020

  20. [28]

    3d-retr: End-to-end single and multi-view 3d reconstruction with transformers, 2021

    Zai Shi, Zhao Meng, Yiran Xing, Yunpu Ma, and Roger Wattenhofer. 3d-retr: End-to-end single and multi-view 3d reconstruction with transformers, 2021

    2021

  21. [29]

    Richter, Ren ´e Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox

    Maxim Tatarchenko, Stephan R. Richter, Ren ´e Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn?, 2019

    2019

  22. [30]

    Multi-view supervision for single-view reconstruction via differentiable ray consistency

    Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2626–2634, 2017

    2017

  23. [31]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS) , pages 6000–6010, 2017

    2017

  24. [32]

    Learning-based multi-view stereo: A survey, 2024

    Fangjinhua Wang, Qingtian Zhu, Di Chang, Quankai Gao, Junlin Han, Tong Zhang, Richard Hartley, and Marc Pollefeys. Learning-based multi-view stereo: A survey, 2024

    2024

  25. [33]

    J. Wang, L. Liu, and Y . Wu. Deep 3d scene understanding: A comprehensive survey. IEEE Access , 8:45678–45692, 2020

    2020

  26. [34]

    Pixel2mesh: Generating 3d mesh models from single rgb images

    Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In European Conference on Computer Vision (ECCV), 2018

    2018

  27. [35]

    Q. Wang, H. Zhang, and M. Lin. Single-view 3d scene reconstruction with high-fidelity shape and texture. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 876–886, 2022

    2022

  28. [36]

    Crm: Single image to 3d textured mesh with convolutional reconstruction model

    Zhengyi Wang, Yikai Wang, Yifei Chen, Chendong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. CoRR, abs/2403.05034, 2024

    2024 arXiv

  29. [37]

    Unique3d: High- quality and efficient 3d mesh generation from a single image, 2024

    Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3d: High- quality and efficient 3d mesh generation from a single image, 2024

    2024

  30. [38]

    3d shapenets: A deep representation for volumetric shapes, 2015

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes, 2015

    2015

  31. [39]

    Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models

    Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191 , 2024

    2024 arXiv

  32. [40]

    K. Xu, Y . Lin, and F. Huang. Holistic 3d scene understanding from a single image with implicit representation. IEEE Trans- actions on Pattern Analysis and Machine Intelligence , 44(1): 102–114, 2022

    2022

  33. [41]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024

    2024

  34. [42]

    Holistic 3d scene under- standing from a single image with implicit representation

    Chenyang Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene under- standing from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 883...

    2021

  35. [43]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes, 2024

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, De- qing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes, 2024

    2024

  36. [44]

    V oxelnet: End-to-end learning for point cloud based 3d object detection

    Yin Zhou and Oncel Tuzel. V oxelnet: End-to-end learning for point cloud based 3d object detection. CoRR, abs/1711.06396, 2017

    2017 arXiv

  37. [45]

    Triplane meets gaussian splatting: Fast and generalizable single-view 3d recon- struction with transformers

    Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d recon- struction with transformers. arXiv preprint arXiv:2312.09147 , 2023

    2023 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.