Pith. sign in

REVIEW 4 major objections 5 minor 77 references

Speed Always Wins: A Survey on Efficient Architectures for Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read RayletDF predicts ray-surface distances from learned raylet segment features and shows single-forward-pass 3D surface reconstruction that generalizes across unseen indoor datasets from point clouds or pre-fit 3D Gaussians.

desk verdict The paper is actually RayletDF, a solid 3D reconstruction paper with a strong generalization claim, but it is wearing an LLM survey's title and its main rival RayDF is a self-built variant, so the headline margins need scrutiny before the claim can be trusted. read the letter →

arxiv 2508.09834 v1 pith:DW2ADTY3 submitted 2025-08-13 cs.CL cs.AIcs.CV

Weigao Sun , Jiaxi Hu , Yucheng Zhou , Jusen Du , Disen Lan , Kexin Wang , Tong Zhu , Xiaoye Qu
show 7 more authors
Yu Zhang Xiaoyu Mo Daizong Liu Yuxuan Liang Wenliang Chen Guoqi Li Yu Cheng
This is my paper · ORCID
classification cs.CLcs.AIcs.CV
keywords modelsarchitecturesefficientlanguagesurveyllmsabovefoundation
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

Most modern 3D reconstruction systems use coordinate-based fields: a network is asked 'is this 3D point on a surface, and how far is it?' and must be evaluated at many points, which is slow. A newer family of ray-based methods instead asks, for each camera ray, 'how far along this ray is the first surface?' That is much faster, but previous ray methods locked their parameters to one specific scene and could not be applied to a new room without retraining.

RayletDF tries to fix the retraining problem. Instead of parameterizing a whole ray, it chops the ray into small segments, called raylets. For each raylet, a feature extractor (a sparse-convolution U-Net) looks at nearby points or 3D Gaussians around that segment, and a small MLP predicts how far the surface is and how confident the prediction is. A blender then combines the raylets' votes. Because the inputs are local geometric features rather than scene-specific embeddings, the same trained network can be pointed at a completely new room and predict surfaces immediately.

The paper trains this network on large indoor datasets (ScanNet, ScanNet++, ARKitScenes) and tests it on held-out rooms and whole unseen datasets (MultiScan). Reported ray-distance errors are roughly half those of the previous ray-based method RayDF, and the method also works from point clouds alone. The caveats: the RayDF numbers come from a version re-implemented by the same research group, the point-cloud test inputs are sampled from clean ground-truth meshes, and the Gaussian input mode still requires an offline per-scene 3D Gaussian fitting step before the 'single forward pass' can run.

Extended reading notes

Core claim

The central claim, stated in Section 4.1: 'When evaluating all methods across new datasets, our method demonstrates superior generalizability on unseen datasets, clearly surpassing all other baselines usually by more than 0.10 meters in accuracy,' and in the abstract: 'our method achieves exceptional generalization ability, successfully recovering 3D surfaces in a single-forward pass across unseen datasets in testing.' If true, a feed-forward network conditioned on local raylet features predicts accurate ray-surface distances on scenes never seen during training, from either point clouds or pre-fit 3D Gaussians.

Load-bearing premise

The headline superiority over existing ray-based methods rests on the authors' own re-implementation of the main rival RayDF, a paper co-authored by two of the current authors. Appendix 5.4 states: 'RayDF: We implement a ray-surface distance field conditioned on the same local geometry features as ours to compare generalization ability fairly in this paper.' If this adaptation under-serves the published per-scene RayDF, the reported cross-dataset margins (e.g., 0.145 vs 0.202 ADE in-domain, and larger gaps on unseen sets) are not evidence against the actual published system. A secondary load-bearing premise is that inputs are representative: points are uniformly sampled from ground-truth meshes and 3D Gaussians are pre-fit per scene after post hoc removal of scenes that 'cannot be pre-estimated by 3DGS' (Appendix 5.1.3, 5.1.4).

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 / 5 minor

Summary. The manuscript presents RayletDF, a feed-forward method for 3D surface reconstruction from either raw point clouds or pre-fitted 3D Gaussians. The method predicts ray-surface distances by defining a 'raylet' as a unit ray segment, extracting local geometry features around sampled raylets, predicting per-raylet distances, and blending multiple raylet predictions. The authors train on ScanNet/ScanNet++ or ARKitScenes and evaluate on those datasets plus MultiScan, reporting distance errors (ADE, RMSE, Abs-Rel, Sq-Rel, threshold accuracy) and mesh metrics. The central claim is that RayletDF generalizes to unseen datasets in a single forward pass, clearly outperforming existing ray-based, Gaussian-splatting, and monocular-depth baselines, often by more than 0.10 m in ADE.

Significance. If the empirical claims hold, RayletDF would be a practically valuable contribution: it offers a single-forward-pass, generalizable surface-reconstruction pipeline from two input modalities, with an extensive evaluation across four real-world datasets, standard-deviation tables, test-time sampling studies, and a speed comparison. The paper is also transparent in acknowledging that depth-estimation baselines are evaluated under favorable scale-alignment conditions and in showing how their performance degrades when aligned to rendered depth (Table 8). However, the significance is currently conditional on resolving a load-bearing baseline concern: the main ray-based competitor, RayDF, is not the published per-scene RayDF but an author re-implementation conditioned on the same local geometry features as RayletDF. Without evidence that this re-implementation is a faithful, strong proxy for the published method, the reported margins (e.g., 0.115 vs. 0.183 ADE in Table 1) do not establish superiority over the actual RayDF system. The data-exclusion steps and the overstatement in Section 4.1 also need attention.

major comments (4)
  1. [§5.4 (Baselines); Tables 1 and 3] The main comparison against RayDF is not against the published RayDF method. Appendix 5.4 states: 'RayDF: We implement a ray-surface distance field conditioned on the same local geometry features as ours to compare generalization ability fairly in this paper.' This is an author-written adaptation, not the NeurIPS 2023 RayDF model, yet Tables 1 and 3 label it 'RayDF [36]' and Section 4.1 uses it as the 'second best method' to support the headline generalization claim. Since two of the current authors are co-authors of the original RayDF, this re-implementation is a particularly sensitive comparison point. The manuscript must either (a) compare against the official RayDF code/checkpoints on the same inputs, (b) calibrate the adapted RayDF to reproduce the published method's per-scene accuracy before using it as a generalization baseline, or (c) relabel the baseline and substantially temper
  2. [§5.1.3 and §5.1.4] The evaluation datasets are filtered by the 3DGS pre-estimation step: for ARKitScenes, 'several rooms which cannot be pre-estimated by 3DGS' are removed, and for MultiScan, 'a few scenes which cannot be pre-estimated by 3DGS' are removed. This is an unquantified, post hoc exclusion that could systematically remove difficult or atypical scenes, directly affecting the measured generalization performance on 'unseen' datasets. The authors should report the number and properties of excluded scenes, justify the exclusion criterion, and show that the main conclusions are unchanged when the excluded scenes are handled by an alternative pipeline or included with a best-effort 3DGS fit.
  3. [§4.1, 'Results & Analysis'] The claim that RayletDF 'clearly surpass[es] all other baselines usually by more than 0.10 meters in accuracy' is not supported by Table 1. For example, when trained on ARKitScenes and tested on ScanNet/ScanNet++, RayletDF achieves ADE 0.175 vs. RayDF 0.227 (gap 0.052) and vs. Pointersect 0.366 (gap 0.191); on MultiScan, RayletDF ADE 0.216 is only 0.050 better than Pointersect 0.266. The 'usually by more than 0.10 meters' wording is an overstatement even if some rows show larger gaps. Please revise the claim to be quantitatively accurate.
  4. [Sections 2–3 (method description)] The manuscript text provided for review does not include the main method sections: Section 3.5 (virtual-ball raylet sampling) is referenced, and the appendix describes the feature extractor and distance-field network, but the core definitions of the raylet distance field, the multi-raylet blender, and the training loss are not fully specified in the submitted text. This is a reproducibility blocker for the central proposal. If the full PDF is missing these sections, the submission is incomplete; if they are absent by design, the method description must be added.
minor comments (5)
  1. [§4.1] Typo: 'Gaussisans' should be 'Gaussians'.
  2. [Table 2] 'Chamber-L1' should be 'Chamfer-L1'.
  3. [Table 8] The column header 'RMS,E' contains a stray comma; also 'RMS' should be 'RMSE'.
  4. [§5.4] The baseline list in Section 4 refers to 'RayDF [36] which is adapted as a generalizable version', but Tables 1 and 3 do not mark this as an adapted version. A clear label such as 'RayDF-adapted' would avoid misreading.
  5. [§5.12] The statement 'RayDF is a per-scene method and we cannot use its pre-trained models' reinforces the need for the calibration described in Major Comment 1; it would be helpful to state this near the main tables as a limitation.

Circularity Check

1 steps flagged · score 4.0 of 10

Generalization claim compared against author re-implementation of RayDF rather than published system.

  1. self citation load bearing [Appendix 5.4 (Baselines); also Section 4.1 Baselines]
    "RayDF: We implement a ray-surface distance field conditioned on the same local geometry features as ours to compare generalization ability fairly in this paper."

    The paper's claimed superiority over RayDF is measured against a baseline that the authors themselves define as sharing 'the same local geometry features as ours' (App. 5.4). Since RayDF [36] is a prior work by two current co-authors (Bo Yang, Jinxi Li), the comparison is not against an external system but against a self-built variant; the 'state-of-the-art' status is imported by citation, and the reported margins (e.g., 0.115 vs 0.183 ADE in Section 4.1) are an artifact of the authors' baseline construction rather than evidence against the published RayDF. The central generalization claim thus reduces in part to a self-referential evaluation, though the method is also compared with independent baselines.

full rationale

The raylet distance field is learned against external ground-truth ray-surface distances on held-out datasets; there is no equation-level circularity in the method's derivation. The main circularity concern is the evaluation protocol: the paper's headline comparison against RayDF [36] uses the authors' own re-implementation ('We implement a ray-surface distance field conditioned on the same local geometry features as ours', App. 5.4) rather than the published per-scene RayDF. This re-implementation shares the proposed method's local geometry feature extractor, so the baseline is a stripped-down version of RayletDF; the reported margins are thus not a test against the actual state-of-the-art ray method but a self-referential ablation. Because two current co-authors (Bo Yang, Jinxi Li) are also co-authors of RayDF, the citation to [36] is not independent support. The central claim nevertheless retains some independent content, as the method also outperforms Pointersect, MVSGaussian, PFGS, and depth estimators on cross-dataset generalization, so the result is not forced by construction. Score 4.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The ledger shows the paper's upstream debt: four hand-chosen sampling parameters (raylet counts, neighbor count, top-K, densification threshold) directly influence the reported numbers, and the method's success rests on two idealized input assumptions (pre-fit per-scene Gaussians; mesh-sampled point clouds with post hoc scene removal). The two invented constructs, the raylet field and virtual-ball sampling, are method-internal and validated only within this paper's own experiments.

free parameters (4)
  • Test-time raylet sample count per ray (N) = 5, 10, 15, and 20 per ray; trained with N=20
    Section 4.3 and Tables 4-6, 12 show ADE improves from 0.194 (N=5) to 0.139 (N=15) on ScanNet/ScanNet++; the paper explicitly markets N as a tunable test-time knob, so the reported headline numbers depend on this choice.
  • Neighbor points sampled per raylet (k) = 5
    Appendix 5.11 states 'each raylet only has 5 neighboring points sampled,' giving the claimed 5x5 = 25 samples per ray efficiency. This hand-chosen count is central to the speed argument and is not ablated.
  • Top-K ray-Gaussian intersections per query ray (K) = not stated explicitly
    Appendix 5.7 describes customized CUDA kernels that select the top K intersection points with the highest alpha blending weight; K is a hand-chosen parameter affecting the conditioning features in the Gaussian input mode.
  • 3DGS densification gradient threshold = 0.0005
    Appendix 5.4 raises the threshold 'to remove floaters' for 3DGS, GOF, and PGSR baselines; this baseline-preparation choice changes the quality of input Gaussians for all compared methods.
assumptions (4)
  • domain assumption A per-scene vanilla 3DGS model pre-fit from RGB images carries enough geometry for raylet-based surface prediction.
    The Gaussian input mode (Section 4.1) depends entirely on offline per-scene 3DGS fitting (37.5 hours for one test split per Appendix 5.11); if the pre-fit Gaussians are poor, the 'single forward pass' claim applies only to the second stage.
  • domain assumption Uniform 10k-point samples from ground-truth meshes are representative of practical sparse point clouds.
    Section 4.2 and Appendix 5.13 sample points from provided scene meshes; the paper claims reconstruction 'from raw point clouds' but tests on idealized mesh-sampled inputs rather than raw LiDAR or SfM output.
  • ad hoc to paper Scenes that cannot be pre-estimated by 3DGS may be excluded without biasing the comparison.
    Appendix 5.1.3 and 5.1.4 remove several ARKitScenes and MultiScan rooms that fail 3DGS pre-estimation; this post hoc, asymmetric exclusion may favor the pipeline.
  • domain assumption Raylet distances from a few sampled local neighbor features suffice to locate the surface along a ray.
    This is the core architectural bet of the paper (Section 3, raylet feature extractor plus MLP predictor); it is supported only by the paper's own experiments, with no external analysis or formal justification.
invented entities (2)
  • Raylet distance field independent evidence
    purpose: A per-unit-segment representation of a ray that enables local geometric conditioning and cross-scene generalization, replacing global ray parameterizations (Plucker, spherical) that limited prior ray-based methods to per-scene training.
    The entity is validated on held-out datasets (unseen ARKitScenes, ScanNet/ScanNet++, MultiScan) in Tables 1-3, which is a falsifiable handle, though the evidence is entirely internal to the paper.
  • Virtual-ball based raylet sampling strategy
    purpose: Samples local neighbor points around each raylet to provide geometric context for the predictor; credited in Section 4.2 with the point-cloud generalization gains.
    No ablation isolates this component; the paper asserts its contribution but does not compare against a variant without virtual-ball sampling, so there is no independent handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Speed Always Wins: A Survey on Efficient Architectures for Large Language Models." pith.science (2026). https://pith.science/paper/DW2ADTY3

@misc{pith2026250809834,
  author       = {Pith},
  title        = {Pith review of: Speed Always Wins: A Survey on Efficient Architectures for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DW2ADTY3}},
  note         = {Machine review of arXiv:2508.09834}
}
read the original abstract

Large Language Models (LLMs) have delivered impressive results in language understanding, generation, reasoning, and pushes the ability boundary of multimodal models. Transformer models, as the foundation of modern LLMs, offer a strong baseline with excellent scaling properties. However, the traditional transformer architecture requires substantial computations and poses significant obstacles for large-scale training and practical deployment. In this survey, we offer a systematic examination of innovative LLM architectures that address the inherent limitations of transformers and boost the efficiency. Starting from language modeling, this survey covers the background and technical details of linear and sparse sequence modeling methods, efficient full attention variants, sparse mixture-of-experts, hybrid model architectures incorporating the above techniques, and emerging diffusion LLMs. Additionally, we discuss applications of these techniques to other modalities and consider their wider implications for developing scalable, resource-aware foundation models. By grouping recent studies into the above category, this survey presents a blueprint of modern efficient LLM architectures, and we hope this could help motivate future research toward more efficient, versatile AI systems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 71 canonical work pages

  1. [36]

    RayDF: Neural Ray-surface Distance Fields with Multi-view Consistency

    Zhuoman Liu, Bo Yang, Yan Luximon, Ajay Kumar, and Jinxi Li. RayDF: Neural Ray-surface Distance Fields with Multi-view Consistency. NeurIPS, 2023. 1, 3, 5, 6, 4, 7

    work page 2023

  2. [1]

    Representing 3D Shapes with Probabilistic Directed Distance Fields

    Tristan Aumentado-Armstrong, Stavros Tsogkas, Sven Dickinson, and Allan Jepson. Representing 3D Shapes with Probabilistic Directed Distance Fields. CVPR, 2022. 1, 3

    work page 2022

  3. [2]

    Fast winding numbers for soups and clouds

    Gavin Barill, Neil G Dickson, Ryan Schmidt, David IW Levin, and Alec Jacobson. Fast winding numbers for soups and clouds. ACM Transactions on Graphics (TOG), 2018. 2

    work page 2018

  4. [3]

    ARK- itScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data

    Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARK- itScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data. NeurIPS,

    work page

  5. [4]

    A Survey of Surface Re- construction from Point Clouds

    M Berger, A Tagliasacchi, L M Seversky, P Alliez, J A Levine, A Sharf, and C T Silva. A Survey of Surface Re- construction from Point Clouds. Computer Graphics Forum,

    work page

  6. [5]

    Richter, and Vladlen Koltun

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth Pro: Sharp Monocular Metric Depth in Less Than a Second. ICLR, 2025. 3, 5

    work page 2025

  7. [6]

    Reid, and John J

    Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, Jose Neira, Ian D. Reid, and John J. Leonard. Past, Present, and Future of Simultaneous Local- ization and Mapping: Towards the Robust-Perception Age. TRO, 2016. 2

    work page 2016

  8. [7]

    Real-Time Neural Light Field on Mobile Devices

    Junli Cao, Huan Wang, Pavlo Chemerys, Vladislav Shakhrai, Ju Hu, Yun Fu, Denys Makoviichuk, Sergey Tulyakov, and Jian Ren. Real-Time Neural Light Field on Mobile Devices. CVPR, 2023. 2

    work page 2023

Show all 77 references
  1. [8]

    Pointersect: Neural Ren- dering with Cloud-Ray Intersection

    Jen-Hao Rick Chang, Wei-Yu Chen, Anurag Ranjan, Kwang Moo Yi, and Oncel Tuzel. Pointersect: Neural Ren- dering with Cloud-Ray Intersection. CVPR, 2023. 1, 5, 6, 4, 7

    2023

  2. [9]

    PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. TVCG,

  3. [10]

    3d re- construction with fast dipole sums

    Hanyu Chen, Bailey Miller, and Ioannis Gkioulekas. 3d re- construction with fast dipole sums. ACM Transactions on Graphics (TOG), 2024. 2

    2024

  4. [11]

    VCR-GauS: View Consis- tent Depth-Normal Regularizer for Gaussian Surface Recon- struction

    Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yun- song Wang, and Gim Hee Lee. VCR-GauS: View Consis- tent Depth-Normal Regularizer for Gaussian Surface Recon- struction. NeurIPS, 2024. 2

    2024

  5. [12]

    Learning Implicit Fields for Generative Shape Modeling

    Zhiqin Chen and Hao Zhang. Learning Implicit Fields for Generative Shape Modeling. CVPR, 2019. 2

    2019

  6. [13]

    Neu- ral Unsigned Distance Fields for Implicit Function Learning

    Julian Chibane, Aymen Mir, and Gerard Pons-Moll. Neu- ral Unsigned Distance Fields for Implicit Function Learning. NeurIPS, 2020. 1, 2

    2020

  7. [14]

    Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese

    Christopher B. Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3D-R2N2: A Unified Approach for Single and Multi-view 3D Object Reconstruction.ECCV,

  8. [15]

    Depth-Regularized Optimization for 3D Gaussian Splatting in Few-Shot Images

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-Regularized Optimization for 3D Gaussian Splatting in Few-Shot Images. CVPRW, 2024. 2

    2024

  9. [16]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scan- Net: Richly-annotated 3D Reconstructions of Indoor Scenes. CVPR, 2017. 5

    2017

  10. [17]

    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. CVPR, 2017. 2

    2017

  11. [18]

    Trim 3D Gaussian Splatting for Accurate Geometry Representation

    Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3D Gaussian Splatting for Accurate Geometry Representation. arXiv:2406.07499, 2024. 2

    2024 arXiv

  12. [19]

    SIGNET: Ef- ficient Neural Representation for Light Fields

    Brandon Yushan Feng and Amitabh Varshney. SIGNET: Ef- ficient Neural Representation for Light Fields. ICCV, 2021. 2

    2021

  13. [20]

    PRIF: Primary Ray-based Implicit Function

    Brandon Y Feng, Yinda Zhang, Danhang Tang, and Ruofei Du. PRIF: Primary Ray-based Implicit Function. ECCV,

  14. [21]

    State of the Art on Neural Rendering

    A Tewari O Fried, J Thies V Sitzmann, S Lombardi K Sunkavalli, and R Martin-brualla T Simon J Saragih. State of the Art on Neural Rendering. Computer Graphics Forum,

  15. [22]

    3D Semantic Segmentation with Submanifold Sparse Convolutional Networks

    Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3D Semantic Segmentation with Submanifold Sparse Convolutional Networks. CVPR, 2018. 3

    2018

  16. [23]

    Grossman and William J

    Jeff P. Grossman and William J. Dally. Point Sample Ren- dering. Rendering Techniques, 1998. 2

    1998

  17. [24]

    SuGaR: Surface- Aligned Gaussian Splatting for Efficient 3D Mesh Recon- struction and High-Quality Mesh Rendering

    Antoine Gu ´edon and Vincent Lepetit. SuGaR: Surface- Aligned Gaussian Splatting for Efficient 3D Mesh Recon- struction and High-Quality Mesh Rendering. CVPR, 2024. 1, 2

    2024

  18. [25]

    Image-based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era

    Xian-feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era. TPAMI, 2019. 2

    2019

  19. [26]

    2D Gaussian Splatting for Geometrically Ac- curate Radiance Fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian Splatting for Geometrically Ac- curate Radiance Fields. SIGGRAPH, 2024. 2

    2024

  20. [27]

    FatesGS: Fast and Accurate Sparse-View Sur- face Reconstruction using Gaussian Splatting with Depth- Feature Consistency

    Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. FatesGS: Fast and Accurate Sparse-View Sur- face Reconstruction using Gaussian Splatting with Depth- Feature Consistency. AAAI, 2025. 2

    2025

  21. [28]

    3DFIRES: Few Image 3D REconstruction for Scenes with Hidden Sur- face

    Linyi Jin, Nilesh Kulkarni, and David Fouhey. 3DFIRES: Few Image 3D REconstruction for Scenes with Hidden Sur- face. CVPR, 2024. 3

    2024

  22. [29]

    Neu- ral 3D Mesh Renderer

    Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neu- ral 3D Mesh Renderer. CVPR, 2018. 2

    2018

  23. [30]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering

    Bernhard Kerbl, Universit ´e C ˆote, Georgios Kopanas, Uni- versit´e Cˆote, Thomas Leimk¨uhler, Max-planck-institut Infor- matik, and G R Aug. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. SIGGRAPH, 2023. 1, 2, 3, 4, 5, 6

    2023

  24. [31]

    Approximate Differ- entiable Rendering with Algebraic Surfaces

    Leonid Keselman and Martial Hebert. Approximate Differ- entiable Rendering with Algebraic Surfaces. ECCV, 2022. 4, 3

    2022

  25. [32]

    Di- rected Ray Distance Functions for 3D Scene Reconstruction

    Nilesh Kulkarni, Justin Johnson, and David F Fouhey. Di- rected Ray Distance Functions for 3D Scene Reconstruction. ECCV, 2022. 1, 3

    2022

  26. [33]

    DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Nor- malization

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing Sparse-View 3D Gaussian Radiance Fields with Global-Local Depth Nor- malization. CVPR, 2024. 1, 2

    2024

  27. [34]

    NeuLF: Efficient Novel View Synthesis with Neural 4D Light Field

    Celong Liu, Zhong Li, Junsong Yuan, and Yi Xu. NeuLF: Efficient Novel View Synthesis with Neural 4D Light Field. Eurographics Symposium on Rendering, 2022. 2

    2022

  28. [35]

    MVSGaussian: Fast Generalizable Gaussian Splatting Re- construction from Multi-View Stereo

    Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. MVSGaussian: Fast Generalizable Gaussian Splatting Re- construction from Multi-View Stereo. ECCV, 2024. 5

    2024

  29. [37]

    MultiScan: Scalable RGBD scanning for 3D environments with articulated objects

    Yongsen Mao, Yiming Zhang, Hanxiao Jiang, Angel X Chang, and Manolis Savva. MultiScan: Scalable RGBD scanning for 3D environments with articulated objects. NeurIPS, 2022. 5

    2022

  30. [38]

    Occupancy Net- works: Learning 3D Reconstruction in Function Space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy Net- works: Learning 3D Reconstruction in Function Space. CVPR, 2019. 1, 2

    2019

  31. [39]

    Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. ECCV, 2020. 1, 2, 6

    2020

  32. [40]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM transactions on graphics (TOG), 2022. 6

    2022

  33. [41]

    Neural Point Light Fields

    Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural Point Light Fields. CVPR, 2022. 2

    2022

  34. [42]

    GigaGS: Scaling up Planar-Based 3D Gaussians for Large Scene Surface Reconstruction

    Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. GigaGS: Scaling up Planar-Based 3D Gaussians for Large Scene Surface Reconstruction. arXiv:2409.06685, 2024. 2

    2024 arXiv

  35. [43]

    A Survey of Structure from Motion

    Onur Ozyesil, Vladislav V oroninski, Ronen Basri, and Amit Singer. A Survey of Structure from Motion. Acta Numerica,

  36. [44]

    DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Represen- tation. CVPR, 2019. 1, 2

    2019

  37. [45]

    Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus H. Gross. Surfels: surface elements as rendering primitives. SIGGRAPH, 2000. 2

    2000

  38. [46]

    Towards Robust Monocu- lar Depth Estimation: Mixing Datasets for Zero-Shot Cross- Dataset Transfer

    Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocu- lar Depth Estimation: Mixing Datasets for Zero-Shot Cross- Dataset Transfer. TPAMI, 2022. 3

    2022

  39. [47]

    Neural Vector Fields for Surface Rep- resentation and Inference

    Edoardo Mello Rella, Ajad Chhatkuli, Ender Konukoglu, and Luc Van Gool. Neural Vector Fields for Surface Rep- resentation and Inference. IJCV, 2024. 2

    2024

  40. [48]

    Object Space EW A Surface Splatting: A Hardware Accelerated Ap- proach to High Quality Point Rendering.Computer Graphics Forum, 2003

    Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object Space EW A Surface Splatting: A Hardware Accelerated Ap- proach to High Quality Point Rendering.Computer Graphics Forum, 2003. 2

    2003

  41. [49]

    Tenenbaum, and Fr ´edo Durand

    Vincent Sitzmann, Semon Rezchikov, William T Free- man, Joshua B. Tenenbaum, and Fr ´edo Durand. Light Field Networks: Neural Scene Representations with Single- Evaluation Rendering. NeurIPS, 2021. 2

    2021

  42. [50]

    GVKF: Gaus- sian V oxel Kernel Functions for Highly Efficient Surface Re- construction in Open Scenes

    Gaochao Song, Chong Cheng, and Hao Wang. GVKF: Gaus- sian V oxel Kernel Functions for Highly Efficient Surface Re- construction in Open Scenes. NeurIPS, 2024. 2

    2024

  43. [51]

    Light Field Neural Rendering

    Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Light Field Neural Rendering. CVPR,

  44. [52]

    MARF: The Medial Atom Ray Field object representation

    Peder Bergebakken Sundt and Theoharis Theoharis. MARF: The Medial Atom Ray Field object representation. Comput- ers & Graphics, 2023. 3

    2023

  45. [53]

    Towards Multi-view Consistency in Neural Ray Fields using Paramet- ric Medial Surfaces

    Peder Bergebakken Sundt and Theoharis Theoharis. Towards Multi-view Consistency in Neural Ray Fields using Paramet- ric Medial Surfaces. Computers & Graphics, 2024. 3

    2024

  46. [54]

    Octree Generating Networks: Efficient Convolutional Archi- tectures for High-resolution 3D Outputs

    Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree Generating Networks: Efficient Convolutional Archi- tectures for High-resolution 3D Outputs. ICCV, 2017. 2

    2017

  47. [55]

    Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik

    Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srini- vasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lom- bardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollh...

    2021

  48. [56]

    GRF: Learning a General Ra- diance Field for 3D Representation and Rendering

    Alex Trevithick and Bo Yang. GRF: Learning a General Ra- diance Field for 3D Representation and Rendering. ICCV,

  49. [57]

    DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing

    Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing. WACV, 2025. 5

    2025

  50. [58]

    RangeUDF: Semantic Surface Reconstruction from 3D Point Clouds

    Bing Wang, Zhengdi Yu, Bo Yang, Jie Qin, Toby Breckon, Ling Shao, Niki Trigoni, and Andrew Markham. RangeUDF: Semantic Surface Reconstruction from 3D Point Clouds. arXiv:2204.09138, 2022. 2

    2022 arXiv

  51. [59]

    PFGS: High Fidelity Point Cloud Rendering via Feature Splatting

    Jiaxu Wang, Ziyi Zhang, Junhao He, and Renjing Xu. PFGS: High Fidelity Point Cloud Rendering via Feature Splatting. ECCV, 2024. 5

    2024

  52. [60]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv:2106.10689, 2021. 6

    2021 arXiv

  53. [61]

    3D Representation Methods: A Survey

    Zhengren Wang. 3D Representation Methods: A Survey. arXiv:2410.06475, 2024. 2

    2024 arXiv

  54. [62]

    GS2Mesh: Surface Reconstruction from Gaussian Splatting via Novel Stereo Views

    Yaniv Wolf, Amit Bracha, and Ron Kimmel. GS2Mesh: Surface Reconstruction from Gaussian Splatting via Novel Stereo Views. ECCV, 2024. 1

    2024

  55. [63]

    Surface Recon- struction from 3D Gaussian Splatting via Local Structural Hints

    Qianyi Wu, Jianmin Zheng, and Jianfei Cai. Surface Recon- struction from 3D Gaussian Splatting via Local Structural Hints. ECCV, 2024. 1, 2

    2024

  56. [64]

    Gaussian- Room: Improving 3D Gaussian Splatting with SDF Guid- ance and Monocular Cues for Indoor Scene Reconstruction

    Haodong Xiang, Xinghui Li, Xiansong Lai, Wanting Zhang, Zhichao Liao, Kai Cheng, and Xueping Liu. Gaussian- Room: Improving 3D Gaussian Splatting with SDF Guid- ance and Monocular Cues for Indoor Scene Reconstruction. arXiv:2405.19671, 2024. 2

    2024 arXiv

  57. [65]

    Neural Fields in Visual Computing and Beyond

    Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp- kin, Vincent Sitzmann, and Srinath Sridhar. Neural Fields in Visual Computing and Beyond. Computer Graphics Forum,

  58. [66]

    DepthSplat: Connecting Gaussian Splatting and Depth

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Polle- feys. DepthSplat: Connecting Gaussian Splatting and Depth. CVPR, 2025. 2

    2025

  59. [67]

    Dense 3D Object Reconstruction from a Sin- gle Depth View

    Bo Yang, Stefano Rosa, Andrew Markham, Niki Trigoni, and Hongkai Wen. Dense 3D Object Reconstruction from a Sin- gle Depth View. TPAMI, 2019. 2

    2019

  60. [68]

    Robust Attentional Aggregation of Deep Feature Sets for Multi-view 3D Reconstruction

    Bo Yang, Sen Wang, Andrew Markham, and Niki Trigoni. Robust Attentional Aggregation of Deep Feature Sets for Multi-view 3D Reconstruction. IJCV, 2020. 2

    2020

  61. [69]

    Depth Any- thing V2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Any- thing V2. NeurIPS, 2024. 3, 5

    2024

  62. [70]

    Neural Vector Fields: Implicit Representation by Explicit Learning

    Xianghui Yang, Guosheng Lin, Zhenghao Chen, and Lup- ing Zhou. Neural Vector Fields: Implicit Representation by Explicit Learning. CVPR, 2023. 1, 2

    2023

  63. [71]

    ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes. ICCV, 2023. 5

    2023

  64. [72]

    Ray-Distance V olume Rendering for Neural Scene Re- construction

    Ruihong Yin, Yunlu Chen, Sezer Karaoglu, and Theo Gev- ers. Ray-Distance V olume Rendering for Neural Scene Re- construction. ECCV, 2024. 3

    2024

  65. [73]

    GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction

    Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xian- gli, and Bo Dai. GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction. NeurIPS, 2024. 2

    2024

  66. [74]

    Gaussian Opacity Fields: Efficient and Compact Surface Reconstruc- tion in Unbounded Scenes

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient and Compact Surface Reconstruc- tion in Unbounded Scenes. SIGGRAPH Asia, 2024. 4, 5, 6, 2

    2024

  67. [75]

    RaDe-GS: Rasterizing Depth in Gaussian Splatting

    Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. RaDe-GS: Rasterizing Depth in Gaussian Splatting. arXiv:2406.01467, 2024

    2024 arXiv

  68. [76]

    Quadratic Gaussian Splatting for Efficient and Detailed Surface Reconstruction

    Ziyu Zhang, Binbin Huang, Hanqing Jiang, Liyang Zhou, Xiaojun Xiang, and Shunhan Shen. Quadratic Gaussian Splatting for Efficient and Detailed Surface Reconstruction. arXiv:2411.16392, 2024. 2

    2024 arXiv

  69. [77]

    Surface splatting

    Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. Surface splatting. SIGGRAPH, 2001. 2 RayletDF: Raylet Distance Fields for Generalizable 3D Surface Reconstruction from Point Clouds or Gaussians Supplementary Material The appendix includes: • Details for ...

    2001

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.