Pith. sign in

REVIEW 4 major objections 6 minor 59 references

HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A hierarchical coarse-to-fine block optimization for 3D Gaussian Splatting reconstructs full-resolution 5K scenes under 24 GB GPU memory while improving quality over full-scene training.

desk verdict Plausible memory-saving 3DGS framework with real ablations, but the view-selection oracle has a blind spot and the verification gaps prevent a clean accept. read the letter →

arxiv 2506.14229 v1 pith:NORANUOF submitted 2025-06-17 cs.CV cs.AI

Changbai Li , Haodong Zhu , Hanlin Chen , Juan Zhang , Tongfei Chen , Shuo Yang , Shuwei Shao , Wenhao Dong
show 1 more author
Baochang Zhang
This is my paper
classification cs.CVcs.AI
keywords 3DGaussianSplattinghigh-resolutionreconstructionmemory-efficientrenderinghierarchicalblockoptimizationSSIM-basedviewselectionpruningsurfacenovelsynthesis
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

The paper claims that hierarchical block-level optimization lets 3D Gaussian Splatting reconstruct full-resolution 5K scenes using only 19 to 23 GB of GPU memory, instead of overflowing a 24 GB card. A coarse global Gaussian model is trained at low resolution, then the scene is contracted into a bounded cube, split into blocks, and each block is refined independently at high resolution using only the views that matter for it. An SSIM-based view-selection rule decides which training cameras see each block, and an importance-pruning step deletes Gaussians that contribute little to rendering. If correct, this makes high-quality novel-view synthesis and surface reconstruction, reported as PSNR 27.91 and SSIM 0.863 on full-resolution Mip-NeRF360 and F1 0.45 on Tanks and Temples, achievable on a single mainstream GPU.

What carries the argument

The central object is the hierarchical block-optimization pipeline with three load-bearing parts: (1) the contracted-space partition, which uses a linear/nonlinear mapping (the contract function of [55]) to normalize unbounded Gaussians into a cubic region so a uniform grid gives balanced blocks; (2) the SSIM-difference view selector (Eq. 3), which compares the full coarse rendering with the rendering after removing the target block's Gaussians and keeps only views whose SSIM loss exceeds the threshold $\epsilon = 0.1$, augmented by boundary poses and binary-search block expansion; (3) Importance-Driven Gaussian Pruning, which scores each Gaussian by opacity times log-volume times a weighted ray-hit count (Eq. 6) and prunes the lowest 20% three times during block refinement. The view-consistent depth-normal regularizer borrowed from VCR-GauS is a supporting component for surface tasks. The whole construction works because the coarse prior anchors every block in the original uncontracted space, so fused blocks stay aligned.

What would settle it

Run HRGS on a scene containing a small patch of fine detail (e.g., printed text on a sign) that is unresolved at the coarse 0.3K resolution, and check whether the block containing that patch receives high-resolution views that see it; the SSIM-difference selector with $\epsilon = 0.1$ should rank those views low and discard them, leaving the patch blurry in the final render. A sharper variant is to corrupt the coarse prior by deleting the patch's Gaussians and measuring how many training views get assigned to that block$-$the count should drop sharply even though the patch is visible from many cameras.

Watch

Extended reading notes

Core claim

The central discovery is that a coarse-to-fine block decomposition guided by a low-resolution global Gaussian prior is sufficient to turn 3DGS into a memory-scaling method: contract the Gaussian set into a bounded cube, partition by a uniform grid, assign to each block the high-resolution views that the coarse model says it affects (via SSIM difference), refine each block in parallel, and fuse. The paper reports that this beats global 3DGS, Mip-Splatting, and VCR-GauS on full-resolution benchmarks while cutting peak GPU memory to 19$-$23 GB and model size roughly in half, and that the two extra mechanisms$-$importance-driven pruning and surface-normal priors$-$are what make the memory savings come without losing fidelity.

Load-bearing premise

The load-bearing premise is that SSIM-difference view selection using the coarse low-resolution global model assigns every block all and only the high-resolution views needed to refine it; if the coarse prior is wrong in a region, informative views show small SSIM differences and are discarded, and the binary-search expansion only adds views by spatial boundaries, so it cannot recover them.

Editorial extensions

If this is right

  • High-resolution, approximately 5K, novel-view synthesis becomes feasible on a single 24 GB GPU, a regime where standard 3DGS runs out of memory.
  • The reported numbers (PSNR 27.91, SSIM 0.863 on Mip-NeRF360; F1 0.45 on Tanks and Temples) position HRGS as a new reference point for high-resolution NVS and surface reconstruction.
  • Importance-Driven Gaussian Pruning roughly halves model size (313.72 MB vs. 621.04 MB without it) while raising SSIM, so pruning acts as both a memory saver and a regularizer.
  • The block count is a practical tuning knob: 4 blocks outperforms 2, 8, and 16 in the ablations, indicating a balance between under-refinement and local overfitting.
  • Because the coarse-to-fine design is resolution-agnostic, the same framework can be applied to even higher resolutions or to 12 GB GPUs by increasing the block count and retuning the view-selection threshold.

Reading between the lines

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

  • The SSIM-difference view selection is essentially a visibility proxy; a purely geometric frustum or ray-cone test against block bounding boxes would be cheaper and independent of coarse-prior quality, and would likely fix the failure mode behind the paper's load-bearing assumption.
  • The contraction-plus-uniform-grid recipe is generic: the same normalization could be applied to other unbounded scene representations, such as NeRF-style volumes or 2DGS surface fields, as a general memory-layering strategy.
  • The fixed pruning schedule (10k/15k/25k iterations, bottom 20%) is likely scene-dependent; scenes with uneven ray coverage could benefit from an adaptive threshold derived from the importance-score distribution.
  • The authors note in the appendix that dynamic-scene extension requires temporal consistency across spatial partitions; a natural follow-up is to make block boundaries time-consistent, but the paper does not attempt that.
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. The paper proposes HRGS, a coarse-to-fine hierarchical framework for 3D Gaussian Splatting (3DGS) that targets high-resolution (≈5K) scene reconstruction under limited GPU memory. A global coarse Gaussian model is first trained on low-resolution images; the scene is then partitioned into spatial blocks (after contracting unbounded space), each block is assigned a subset of training views via an SSIM-difference criterion plus a boundary-based criterion, and each block is refined in parallel with high-resolution data. An Importance-Driven Gaussian Pruning (IDGP) strategy prunes low-contribution Gaussians during refinement, and a depth-normal regularizer from VCR-GauS is used for surface reconstruction. Experiments on Mip-NeRF 360, Tanks and Temples, and Replica report state-of-the-art or competitive PSNR/SSIM/LPIPS and F1 scores, with GPU memory of 19–23 GB and smaller model sizes than 3DGS and CityGS.

Significance. If the results are reproducible and the comparisons fair, the paper makes a useful contribution: it shows that a coarse-to-fine block decomposition, combined with importance-based pruning, can enable 5K-scale Gaussian splatting on a single 24 GB GPU while improving rendering quality over full-scene 3DGS. The full-dataset ablations in the supplementary (Tab. C) and the comparison against CityGS (Tab. A, B) are valuable, and the IDGP ablation shows a substantial model-size/memory reduction. However, the headline claims rest on two pillars that are not fully secured: the SSIM-based view-selection oracle in Eq. (3) and the comparability of the reported baselines. The paper is a solid candidate after these concerns are addressed with additional experiments and reporting.

major comments (4)
  1. [§3.2, Eq. (3)–(5)] The SSIM-difference view-selection criterion is a measure of a view's present contribution to the coarse model, not of the block's information need. If the coarse prior has no Gaussians (or misplaced Gaussians) in a region, removing the block's Gaussians changes the coarse rendering by almost nothing, so the very views that could fix that region fail the epsilon=0.1 threshold and are discarded. Eq. (4) adds only poses whose camera centers lie inside the block's contracted bounds, which is not the same as poses that see the block; outdoor blocks with all cameras outside that volume can receive P2 = empty. The binary-search expansion terminates on the Gaussian count K_j, not on view coverage. No per-block selected-view counts, coverage statistics, or robustness to coarse-model corruption are reported in Tables 1–3 or the supplementary. This is load-bearing because the data partitioning determines the training signal for every block. Please report per-block view coverage and add an experiment that corrupts or removes parts of the coarse prior to test whether informative views are still recovered; or compare against an oracle visibility-based view selection.
  2. [Table 1 and Table E] The full-resolution baseline numbers call for justification. 3DGS is reported at PSNR 19.59/SSIM 0.619 on Mip-NeRF 360 full resolution, which is far below the values commonly reported for 3DGS on this dataset even when trained at default resolution, and below the same method's own 1/8-resolution result of 29.19/0.880 in Table E. The paper says methods exceeding the A5000 memory were run on A800, but it does not state the exact training resolution, number of iterations, or hyperparameters used for each baseline, nor whether the baselines were re-trained from scratch or taken from official checkpoints. If 3DGS and Mip-Splatting were run at 5K with default 1K-oriented settings, they may be artificially handicapped, undermining the claim of state-of-the-art performance. Please specify the protocol for each baseline, report the GPU memory each method actually consumed, and consider including a matched-memory comparison (e.g., 3DGS trained with reduced resolution or in blocks) to isolate the effect of the proposed framework.
  3. [Tables 1–3 and supplementary Tabs. C–E] All quantitative results are single runs with no error bars or number of seeds. Some of the claimed improvements are small (e.g., TNT mean F1 0.45 vs. 0.40 for VCR-GauS; Replica F1 74.87 vs. 64.36 for 2DGS; IDGP ablation PSNR 26.39 vs. 26.41), and the direction of the IDGP effect changes between PSNR and SSIM. Without variance estimates, the reader cannot tell whether the differences are significant. Please report at least three seeds for the main tables, or, if full-dataset runs are too expensive, provide per-scene standard deviations and a statistical test on the mean improvements.
  4. [§3.2 and §4.1] The description of parallel block refinement is ambiguous regarding GPU memory. The text says blocks are 'refined in parallel' on a single GPU and reports a single memory number (19–23 GB). It is not stated whether this is peak memory for the entire pipeline (coarse training plus all blocks), whether blocks are trained simultaneously on one GPU or sequentially, or whether the reported memory is for the refinement stage only. Since memory efficiency is the paper's central claim, please define exactly how memory is measured (peak vs. average, which stage) and describe the parallelization schedule (threads, multi-process, or sequential loops).
minor comments (6)
  1. [Throughout] Cross-references are systematically incorrect: 'as shown in Tab. C' (Sec. 4.1, NVS paragraph) refers to Table 1 in the main text while the supplementary also has a Table C; 'Tab. E' and 'Fig. D' are used for main-text tables; 'Tab D' appears instead of Table D; and 'IDPG' is used interchangeably with 'IDGP' in several places (e.g., Table D caption). Please unify the numbering and fix the typos.
  2. [Eq. (10)] The loss L_s is referenced in Eq. (10) as 'introduced to simplify depth computation, as described in [13]', but is never defined in the paper. Define L_s explicitly or remove it from the equation.
  3. [Supplementary Tab. B] In the 'room' row of Table B, the total model size is 221.97 MB but cell0 is listed as 316.31 MB, which exceeds the total. This is either a typo or a serious reporting error; please correct it and check the other rows for consistency.
  4. [§4.1] The coarse stage is said to train at 'low resolution (0.3K)'. Since the paper targets 5K images, please specify the exact pixel dimensions or downsampling factor, because 0.3K is ambiguous (300 pixels wide vs. 0.3× the original resolution).
  5. [§4.1 and §3.2 (IDGP)] The IDGP pruning step removes the lowest 20% of Gaussians at fixed iterations, but it is not stated whether pruned Gaussians can be re-created by the standard 3DGS adaptive density control during the remaining iterations. If they can, the reported memory reduction may differ from the intended mechanism; if they cannot, this should be stated explicitly.
  6. [General] The paper does not state whether code and trained models will be released. Given the importance of reproducing memory measurements, a statement of code availability would be helpful.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the reported gains rest on external benchmarks and a heuristic block-partitioning scheme, not on a fitted prediction.

full rationale

The paper's derivation chain is empirical rather than circular. A coarse low-resolution global Gaussian model is first trained with the standard 3DGS photometric loss; the scene is then partitioned by the deterministic contraction and grid rules of Eqs. (2)-(5); each block is refined on high-resolution views using the same photometric loss plus external normal priors; and the final PSNR/SSIM/LPIPS/F1 numbers are measured on standard external benchmarks. No parameter is fitted to the reported metrics and then renamed as a prediction. The SSIM-based view selection in Eq. (3) is a heuristic data-partitioning rule driven by the coarse prior, not a derivation of the final reconstruction quality; the concern that it may discard informative views when the coarse prior is inaccurate is a robustness/correctness limitation, not an equation-level circularity. The only author-overlap citation is VCR-GauS [13] for the depth-normal regularizer, whose lead author is a co-author here; however, that regularizer is an independently published component and is not the source of the paper's central memory-efficiency claim. The manuscript's own limitation section discusses dynamic-scene extension, not a circular step. Overall, the central claim is not forced by definition or by self-citation.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on six domain assumptions about initialization, contraction, view assignment, pruning scoring, normal priors, and block fusion. None are proven; the ablations provide empirical support for the data-division and pruning components, but not for the accuracy of the coarse prior or the normal priors. There are seven hand-set hyperparameters, most without sensitivity analysis.

free parameters (7)
  • SSIM threshold epsilon = 0.1
    Used in Eq. 3 to assign camera poses to blocks; chosen by hand, no sensitivity analysis reported.
  • IDGP prune ratio = 20%
    The lowest-importance 20% of Gaussians are removed at each pruning step; no ablation on this ratio is reported.
  • number of blocks n = 4
    Selected by ablation on 2, 4, 8, and 16 blocks on TNT; used for all datasets.
  • coarse resolution scale = 0.3K
    Coarse global Gaussian is trained at 0.3K resolution for 30,000 iterations; hand-chosen.
  • loss weights lambda_1, lambda_2, lambda_3 = 1, 0.01, 0.015
    Weights for surface, normal, and D-normal losses; fixed without sensitivity study.
  • IDGP pruning schedule = 10k, 15k, 25k iterations
    Pruning applied at these iterations during each block's 30,000-iteration refinement; hand-chosen.
  • internal region fraction = central one-third
    Boundary between linear and nonlinear contraction in Eq. 2 is set to the central one-third of the scene; arbitrary design choice.
assumptions (6)
  • domain assumption The COLMAP sparse point cloud provides a sufficiently dense and accurate initialization for coarse Gaussian training.
    Invoked in Section 3.2, where the paper trains the COLMAP points using all observations at low resolution; errors here propagate to block partitioning and view assignment.
  • domain assumption ScanNeRF contraction (Eq. 2) maps the unbounded scene into a bounded cube while preserving spatial locality.
    The uniform-grid block division depends on contracted coordinates being spatially meaningful; adopted from [55] without validation on these scenes.
  • domain assumption SSIM differences between renderings with and without a block identify all relevant training views for that block.
    Eq. 3 uses this to build P1_j; this is the weakest premise because coarse-model errors can cause informative views to be discarded.
  • domain assumption The importance score S_i = alpha_i * ln(1+v_i) * H_i is a reliable proxy for a Gaussian's contribution to reconstruction quality.
    Used by IDGP in Eq. 6; no formal or empirical link between this score and final rendering error is established.
  • domain assumption Pretrained monocular normal predictors (DSINE for outdoor, GeoWizard for indoor) provide supervision accurate enough to improve surfaces.
    Eqs. 7 to 9 penalize deviations from predicted normals; if the predictions are biased, the reconstructed surface inherits the bias.
  • domain assumption Independent block refinement initialized from the coarse prior yields a globally consistent fused model.
    The paper asserts seamless fusion in Section 3.2 but provides no theoretical or quantitative test of inter-block consistency beyond aggregate metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction." pith.science (2026). https://pith.science/paper/NORANUOF

@misc{pith2026250614229,
  author       = {Pith},
  title        = {Pith review of: HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NORANUOF}},
  note         = {Machine review of arXiv:2506.14229}
}
read the original abstract

3D Gaussian Splatting (3DGS) has made significant strides in real-time 3D scene reconstruction, but faces memory scalability issues in high-resolution scenarios. To address this, we propose Hierarchical Gaussian Splatting (HRGS), a memory-efficient framework with hierarchical block-level optimization. First, we generate a global, coarse Gaussian representation from low-resolution data. Then, we partition the scene into multiple blocks, refining each block with high-resolution data. The partitioning involves two steps: Gaussian partitioning, where irregular scenes are normalized into a bounded cubic space with a uniform grid for task distribution, and training data partitioning, where only relevant observations are retained for each block. By guiding block refinement with the coarse Gaussian prior, we ensure seamless Gaussian fusion across adjacent blocks. To reduce computational demands, we introduce Importance-Driven Gaussian Pruning (IDGP), which computes importance scores for each Gaussian and removes those with minimal contribution, speeding up convergence and reducing memory usage. Additionally, we incorporate normal priors from a pretrained model to enhance surface reconstruction quality. Our method enables high-quality, high-resolution 3D scene reconstruction even under memory constraints. Extensive experiments on three benchmarks show that HRGS achieves state-of-the-art performance in high-resolution novel view synthesis (NVS) and surface reconstruction tasks.

Figures

Figures reproduced from arXiv: 2506.14229 by the authors.

Figure 1
Figure 1. (a) High-resolution (∼ 5K) renderings of the “bicycle” scene from the Mip-NeRF 360 dataset, with results from 3DGS, Mip-Splatting, and our method. Red dashed boxes highlight key details. (b) Performance on this scene: our method achieves the highest PSNR (25.6 dB) with significantly lower GPU memory (23 GB) and model size (567 MB) than 3DGS and Mip-Splatting. (c) Average resource usage across the full dataset shows … view at source ↗
Figure 2
Figure 2. Illustrative diagram of the hierarchical block optimization framework. We first derive a global coarse Gaussian representation using low-resolution data, which is then contracted into a bounded cubic region. Subsequently, the contracted Gaussian primitives are partitioned into blocks, each paired with corresponding data. Leveraging the global coarse Gaussian as initialization, we parallelly refine each block in the … view at source ↗
Figure 3
Figure 3. Qualitative Comparison on the Mip-NeRF 360 Dataset. Three representative scenes demonstrate that our method more faithfully preserves fine-scale structures and achieves superior visual fidelity compared to 3DGS and Mip-Splatting. scale factor of pi along the d-th spatial axis, and apply logarithmic compression vei = ln 1 + vi  . Finally, we assign each primitive an importance score with its opacity αi: Si = αi vei … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative Comparison on TNT dataset. Reconstructions from left to right—SuGar, NeuS, 2DGS, and VCR-Gaus—demonstrate that our method delivers more complete surface ge￾ometry, enhanced smoothness in planar regions, and superior preservation of fine structural details, …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 55 canonical work pages

  1. [1]

    Deep multi-view learning methods: A review.Neurocomputing, 448:106–129, 2021

    work page 2021

  2. [2]

    Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In CVPR, 2024

    work page 2024

  3. [3]

    Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition 2024, Seattle, WA, USA, June 16-22, 2024, pages 9535–9545, 2024

    work page 2024

  4. [4]

    Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B. Goldman. Patchmatch: a randomized correspondence algorithm for structural image editing.ACM Trans. Graph., 28(3):24, 2009

    work page 2009

  5. [5]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5460–5469. IEEE, 2022

    work page 2022

  6. [6]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022

    2022

  7. [7]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti- aliased grid-based neural radiance fields. InIEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19640–19648. IEEE, 2023

    work page 2023

  8. [8]

    Patchmatch stereo - stereo matching with slanted support windows

    Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo - stereo matching with slanted support windows. InBritish Machine Vision Conference, BMVC 2011, Dundee, UK, August 29 - September 2, 2011. Proceedings, pages 1–11. BMV A Press, 2011

    work page 2011

Show all 59 references
  1. [9]

    Patchmatch stereo - stereo matching with slanted support windows

    Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo - stereo matching with slanted support windows. In Jesse Hoey, Stephen J. McKenna, and Emanuele Trucco, editors,British Machine Vision Conference, BMVC 2011, Dundee, UK, August 29 - September 2, 2011. Pro...

    2011

  2. [10]

    A probabilistic framework for space carving

    Adrian Broadhurst, Tom Drummond, and Roberto Cipolla. A probabilistic framework for space carving. InProceedings of the Eighth International Conference On Computer Vision (ICCV-01), V ancouver , British Columbia, Canada, July 7-14, 2001 - V olume 1, pages 388–393. IEEE Compute...

    2001

  3. [11]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InComputer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, volume 13692, pages 333–350. Springer, 2022

    2022

  4. [12]

    Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance.CoRR, abs/2312.00846, 2023

    Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance.CoRR, abs/2312.00846, 2023

    2023 arXiv

  5. [13]

    Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction

    Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng...

    2024

  6. [14]

    Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction

    Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing S...

    2024

  7. [15]

    Learning implicit fields for generative shape modeling

    Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. InIEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, June 16-20, 2019, pages 5939–5948. Computer Vision Foundation / IEEE, 2019

    2019

  8. [16]

    Drebin, Loren Carpenter, and Pat Hanrahan

    Robert A. Drebin, Loren Carpenter, and Pat Hanrahan. V olume rendering. InProceedings of the 15th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’88, page 65–74, 1988

    1988

  9. [17]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5491–...

    2022

  10. [18]

    Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction

    Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Proces...

    2022

  11. [19]

    Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image

    Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, an...

    2024

  12. [20]

    Accurate, dense, and robust multiview stereopsis.IEEE Trans

    Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis.IEEE Trans. Pattern Anal. Mach. Intell., 32(8):1362–1376, 2010

    2010

  13. [21]

    Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F

    Steven J. Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F. Cohen. The lumigraph. InPro- ceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH 1996, New Orleans, LA, USA, August 4-9, 1996, pages 43–54. ACM, 1996

    1996

  14. [22]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh re- construction and high-quality mesh rendering

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh re- construction and high-quality mesh rendering. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 16-22, 2024, pages 5354–5363. IEEE, 2024

    2024

  15. [23]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In Andres Burbano, Denis Zorin, and Wojciech Jarosz, editors, ACM SIGGRAPH 2024 Conference Papers, Denver , CO, USA, 27 July 2024- 1 August 2...

    2024

  16. [24]

    Tri-perspective view for vision-based 3d semantic occupancy prediction

    Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 9223–9232, June 2023

    2023

  17. [25]

    Distributed poisson surface reconstruction.Comput

    Michael Kazhdan and Hugues Hoppe. Distributed poisson surface reconstruction.Comput. Graph. F orum, 42(6), 2023

    2023

  18. [26]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139:1–139:14, 2023

    2023

  19. [27]

    Tanks and temples: benchmarking large-scale scene reconstruction.ACM Trans

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction.ACM Trans. Graph., 36(4):78:1–78:13, 2017

    2017

  20. [28]

    Kutulakos and Steven M

    Kiriakos N. Kutulakos and Steven M. Seitz. A theory of shape by space carving.Int. J. Comput. Vis., 38(3):199–218, 2000

    2000

  21. [29]

    Efficient ray tracing of volume data.ACM Trans

    Marc Levoy. Efficient ray tracing of volume data.ACM Trans. Graph., 9(3):245–261, 1990

    1990

  22. [30]

    Light field rendering

    Marc Levoy and Pat Hanrahan. Light field rendering. InProceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’96, page 31–42, 1996

    1996

  23. [31]

    A quasi-dense approach to surface reconstruction from uncalibrated images.IEEE Trans

    Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images.IEEE Trans. Pattern Anal. Mach. Intell., 27(3):418–433, 2005

    2005

  24. [32]

    Binary search algorithm.WikiJournal of Science, 2(1):1–13, 2019

    Anthony Lin. Binary search algorithm.WikiJournal of Science, 2(1):1–13, 2019

    2019

  25. [33]

    Vastgaussian: Vast 3d gaussians for large scene recon- struction

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene recon- struction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, Seat...

    2024

  26. [34]

    Neural sparse voxel fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. InAdvances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020

    2020

  27. [35]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians

    Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. InComputer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, ...

    2024

  28. [36]

    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 synthesis. InComputer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceeding...

    2020

  29. [37]

    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 synthesis.Commun. ACM, 65(1):99–106, 2022

    2022

  30. [40]

    Instant neural graphics primitives with a multiresolution hash encoding.ACM Trans

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding.ACM Trans. Graph., 41(4):102:1–102:15, 2022

    2022

  31. [41]

    Florence, Julian Straub, Richard A

    Jeong Joon Park, Peter R. Florence, Julian Straub, Richard A. Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. InIEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, June 16-20, 2019, p...

    2019

  32. [42]

    Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T

    Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: memory-efficient radiance fields for real-time view synthesis in unbounded scenes.ACM Trans. Graph., 42(4):89:1–89:12, 2023

    2023

  33. [43]

    Schonberger and Jan-Michael Frahm

    Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, June 2016

    2016

  34. [44]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Computer Vision and Pattern Recognition (CVPR), 2016. 12

    2016

  35. [46]

    Sch ¨onberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys

    Johannes L. Sch ¨onberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view se- lection for unstructured multi-view stereo. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors,Computer Vision - ECCV 2016 - 14th European Conference, Amsterdam, ...

    2016

  36. [47]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes Lutz Sch ¨onberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. InEuropean Conference on Computer Vision (ECCV), 2016

    2016

  37. [48]

    Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski

    Steven M. Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2006), 17-22 June 2006, N...

    2006

  38. [49]

    Seitz and Charles R

    Steven M. Seitz and Charles R. Dyer. Photorealistic scene reconstruction by voxel coloring.Int. J. Comput. Vis., 35(2):151–173, 1999

    1999

  39. [50]

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Yuheng Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, T...

    1906 arXiv

  40. [51]

    Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 18-24, 2022, pages 5449–5459. IEEE

    2022

  41. [52]

    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. InCVPR, 2024

    2024

  42. [53]

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

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learn- ing neural implicit surfaces by volume rendering for multi-view reconstruction. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaug...

    2021

  43. [54]

    Wang, E.P

    Z. Wang, E.P. Simoncelli, and A.C. Bovik. Multiscale structural similarity for image quality assessment. InThe Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, pages 1398–1402 V ol.2, 2003

    2003

  44. [55]

    Scanerf: Scalable bundle-adjusting neural radiance fields for large-scale scene rendering.ACM Trans

    Xiuchao Wu, Jiamin Xu, Xin Zhang, Hujun Bao, Qixing Huang, Yujun Shen, James Tompkin, and Weiwei Xu. Scanerf: Scalable bundle-adjusting neural radiance fields for large-scale scene rendering.ACM Trans. Graph., 42(6):261:1–261:18, 2023

    2023

  45. [56]

    Srinivasan, Richard Szeliski, Jonathan T

    Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. InACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH 2023, Los Angeles, CA, USA, ...

    2023

  46. [57]

    Differentiable surface splatting for point-based geometry processing.ACM Transactions on Graphics, 38(6):1–14, 2019

    Wang Yifan, Felice Serena, Shihao Wu, Cengiz ¨Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing.ACM Transactions on Graphics, 38(6):1–14, 2019

    2019

  47. [58]

    Enforcing geometric constraints of virtual normal for depth prediction

    Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. InICCV, pages 5684–5693, 2019

    2019

  48. [59]

    Mip-splatting: Alias-free 3d gaussian splatting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19447–19456. IEEE, 2024

    2024

  49. [60]

    Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction

    Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Info...

    2022

  50. [61]

    Open3d: A modern library for 3d data processing.CoRR, abs/1801.09847, 2018

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing.CoRR, abs/1801.09847, 2018

    2018 arXiv

  51. [62]

    SO Ass.” refers to SSIM-based assignment, while “BO Ass

    Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EW A volume splatting. In Thomas Ertl, Kenneth I. Joy, and Amitabh Varshney, editors,12th IEEE Visualization Conference, IEEE Vis 2001, San Diego, CA, USA, October 24-26, 2001, Proceedings, pages 29–36....

    2001

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.