Pith. sign in

REVIEW 3 major objections 7 minor 38 references

Hybrid Mesh-Gaussian Representation for Efficient Indoor Scene Reconstruction

T0 review · 3 major / 7 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 Textured meshes can replace most Gaussians on flat indoor surfaces, cutting primitive count by 18-35% with no measurable quality drop.

desk verdict The hybrid mesh-Gaussian idea is genuinely useful and the Gaussian-count reductions are real, but the headline FPS advantage is unproven because the timing excludes the mesh render pass. read the letter →

arxiv 2506.06988 v1 pith:C4SJ7UEF submitted 2025-06-08 cs.CV

Binxiao Huang , Zhihao Li , Shiyong Liu , Xiao Tang , Jiajun Tang , Jiaqi Lin , Yuxin Cheng , Zhenyu Chen
show 2 more authors
Xiaofei Wu Ngai Wong
This is my paper
classification cs.CV
keywords hybridmesh-Gaussianrepresentation3DGaussiansplattingtexturedmeshindoorscenereconstructionpruningtransmittance-awaresupervisionreal-timerenderingnovelviewsynthesis
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 proposes a hybrid scene representation for indoor reconstruction that splits labor between two complementary primitives: textured triangle meshes own flat, texture-rich surfaces such as walls and floors, while 3D Gaussians keep modeling geometrically intricate regions such as furniture edges and thin structures. The authors argue that vanilla 3DGS wastes thousands of Gaussian splats on flat areas just to encode high-frequency color variation, and that a texture map expresses the same content far more compactly. To make the hybrid work, they prune the extracted mesh in geometrically complex areas using normal-variation, adjacency-angle, and triangle-area metrics, then jointly optimize Gaussians and mesh with a transmittance-aware texture loss. On ScanNet++ and Deep Blending they report 18-35% fewer Gaussians than 3DGS with PSNR changes under 0.07 dB and higher rasterization FPS, plus sharper texture in flat regions. The load-bearing idea is that texture maps can substitute for many splats without sacrificing image quality.

What carries the argument

The mechanism is a mesh-pruning and joint-optimization pipeline. Pruning removes triangles in geometrically unreliable areas using three criteria: high total variation of a StableNormal prior normal map, adjacent-face angles above 45 degrees, and the smallest area percent of triangles, followed by removal of isolated connected components, hole closing, and subdivision smoothing. The surviving mesh is UV-mapped with Xatlas and textured with Nvdiffrast, then rendered as an opaque background with fixed depth while Gaussians rasterize in front; the final color is $I_h = I_{gs} + T \cdot I_m$ when a triangle covers the pixel. A transmittance-aware mask $M_T = 1/(1+e^{-k(T-0.5)})$ gates the texture loss so that pixels where mesh transmittance is high are pushed to match the ground truth, while pixels where foreground Gaussians dominate are protected from ghosting, and a warm-up phase lets Gaussians populate missing objects before the texture loss switches on.

What would settle it

Render the mesh-only contribution (before Gaussians are composited) on held-out novel views of a flat textured wall with a slight undulation or a wall-to-ceiling corner where the extracted mesh has known error; if the texture map shows ghosting, blur, or seams at the Gaussian-mesh boundary, then the claim that textured meshes can substitute for Gaussians without quality loss fails in exactly the regions the method targets. A cleaner test: on the Replica upper-bound setup, perturb the ideal mesh vertices by a small noise (e.g., 1 cm) and measure how much of the reported PSNR gain disappears.

Watch

Extended reading notes

Core claim

The central claim is that a textured mesh treated as an opaque background at a known depth can absorb the representational burden of flat, texture-rich indoor regions, allowing a hybrid mesh-Gaussian model to render at comparable quality with substantially fewer Gaussian primitives and higher frame rate. The authors establish this by first extracting a mesh from PGSR, pruning it with heuristic geometric metrics to remove regions where the mesh is unreliable, baking an optimizable texture map, and then jointly training the Gaussians and the mesh. A warm-up period plus a transmittance-aware mask prevents the texture loss from painting foreground object colors onto the background mesh. Experiments on ten real indoor scenes show the hybrid uses 0.742M versus 0.911M Gaussians on ScanNet++ at essentially unchanged PSNR (24.28 vs 24.22 dB) and lifts rasterizer FPS from 211 to 231; on Deep Blending the count drops from 2.634M to 1.698M with FPS rising from 346 to 498. On the synthetic Replica dataset, where mesh geometry is exact, the same pipeline cuts Gaussians from 1.667M to 0.295M and raises FPS from 122 to 446, indicating the ceiling of the approach.

Load-bearing premise

The retained mesh geometry is accurate enough in flat regions that a single texture map baked from training views reproduces correct colors from every viewpoint, because the optimizer can adjust texture colors but cannot fix wrong vertex positions.

Editorial extensions

If this is right

  • Indoor scenes with large planar textured surfaces (walls, floors, ceilings) can be rendered with 18-35% fewer Gaussian primitives at essentially unchanged PSNR, and the freed compute converts directly into higher FPS in the CUDA rasterizer.
  • Compositing a textured mesh as an opaque background at fixed depth is enough to suppress Gaussian densification in flat regions: the correct color from the mesh lowers the photometric gradient that would otherwise trigger clone and split operations.
  • Combining the mesh prior with existing Gaussian-pruning methods compounds the savings: on Reduced GS, the hybrid cuts Gaussians from 1.333M to 0.668M on Deep Blending with negligible metric change.
  • If a high-quality mesh is available, the same pipeline approaches 0.295M Gaussians at 446 FPS on Replica, suggesting that better mesh extraction is a direct path to further compression.
  • Texture maps store high-frequency appearance more cheaply than splats, so the method trades a 3×2048×2048 texture map for hundreds of thousands of Gaussians, reducing per-scene memory in the splat budget.

Reading between the lines

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

  • Beyond the paper: the heuristic pruning thresholds (α normal, α area, α group) could be replaced by a learned importance predictor that scores triangles from training-view normal and depth errors, potentially pruning more aggressively without manual tuning.
  • Beyond the paper: the transmittance-aware mask is a generic mechanism usable in any hybrid renderer that composites an opaque background behind transparent splats, including outdoor or object-centric scenes with planar priors, not just indoor walls.
  • Beyond the paper: because FPS is measured on the CUDA rasterizer only, the reported speedup reflects splat-count reduction; end-to-end performance on a lightweight GPU would also include mesh rasterization and texture sampling, which may narrow the gap for small splat budgets.
  • Beyond the paper: the authors' own limitation statement (less effective in geometry-dominated scenes without a good mesh) implies that porting the same architecture to stronger mesh-extraction backbones than PGSR could convert more of the scene to textured mesh and yield further compression.
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

3 major / 7 minor

Summary. The paper proposes a hybrid representation for indoor scene reconstruction that combines 3D Gaussian splatting with textured triangle meshes. A mesh extracted with PGSR is pruned using normal-variation, adjacent-angle, triangle-area, and connectivity thresholds, textured via Xatlas and Nvdiffrast, and then jointly optimized with the Gaussians under a transmittance-aware mask and a warm-up schedule. Experiments on Deep Blending and ScanNet++ compare the method against 3DGS, Mip-splatting, and Reduced GS, reporting PSNR, SSIM, LPIPS, Gaussian counts, and FPS. The paper claims that the hybrid representation maintains comparable rendering quality while requiring fewer Gaussians and achieving superior FPS.

Significance. The direction is sensible: using textured meshes for flat texture-rich regions and reserving Gaussians for complex geometry is a plausible way to reduce primitive count without giving up high-frequency appearance. The Gaussian-count reductions are consistent across ten scenes and three baselines, and the paper provides useful ablations (mesh pruning, transmittance mask, texture-loss weight, texture initialization) and per-scene tables in the supplementary material. However, the headline efficiency claim rests on rasterizer-only FPS measurements that omit the mesh rendering and compositing steps required by the hybrid, and the quality-parity claim is weakened by a systematic LPIPS degradation. The core mechanism also depends on the retained mesh being geometrically accurate, which the paper itself concedes is not always the case. With end-to-end timing and a more nuanced quality discussion, this could be a solid contribution.

major comments (3)
  1. [§4.1, §3.4, Table 1] The headline claim of "superior FPS" is not supported by the reported measurements. Section 4.1 states that FPS numbers "isolate the runtime of the CUDA rasterizer routine only and exclude any graphics API overheads." For the hybrid method, however, §3.4 requires, for every frame, rendering the mesh with Nvdiffrast to obtain I_m and D_m, computing the per-pixel triangle availability and the transmittance mask, running the GS rasterizer, and compositing I_h = I_gs + T × I_m. None of the mesh-related work is included in the reported FPS, whereas a 3DGS frame consists only of the GS rasterizer. The comparison is therefore not apples-to-apples; on Deep Blending, the hybrid is credited with 498 FPS versus 346 FPS for 3DGS in Table 1 while retaining 1.698M Gaussians, yet the mesh rendering and compositing costs could erase that margin. The authors should report end-to-end FPS using the same rendering path for both methods, or at least give per-component timings so the overhead of the mesh stage is visible.
  2. [Table 1, Table 8, §4.2] The claim of "comparable rendering quality" is not fully supported by the LPIPS results. On Deep Blending, LPIPS worsens for all four scenes in Table 8 for the 3DGS + Mesh variant (playroom 0.249→0.255, drjohnson 0.250→0.262, bedroom 0.282→0.290, creepyattic 0.231→0.243), and the mean in Table 1 worsens from 0.253 to 0.262; on ScanNet++, LPIPS also worsens in every scene reported for this variant. PSNR changes are small in the 3DGS comparison (within 0.07 dB on the Table 1 means), but the direction is mixed and some per-scene PSNR drops are larger (bedroom 28.93→28.69, creepyattic 30.33→30.11 in Table 7). Because no error bars or variance estimates are provided, it is not possible to determine whether these differences are significant. The authors should either qualify the quality claim to the metrics and scene subsets where it holds, or provide statistical evidence and explain why the consistent LPIPS degradation is acceptable.
  3. [Supplementary B, Limitations] The core mechanism assumes that the retained mesh geometry is accurate enough in flat textured regions that baking a single texture from all training views yields view-consistent colors. The paper's own supplementary section B states that "even after pruning, there remain regions within the mesh that possess geometric errors," and the Limitations section states that vertex coordinates are not optimized. Since the joint optimization can only correct texture colors, not wrong vertex positions, these residual errors can produce blurry or view-inconsistent textures in the regions the mesh is meant to represent. The authors should quantify the geometric accuracy of the retained mesh (for example, depth error maps against the optimized Gaussians, or the fraction of pixels with large |D_m - D_gs|) and show that quality parity holds in mesh-dominated regions. Without such evidence, the mechanism by which the hybrid maintains quality is not verified.
minor comments (7)
  1. [§1] The Introduction cites Mildenhall et al. 2020 for 3D Gaussian splatting; the correct reference is Kerbl et al. 2023.
  2. [Table 1 caption] The caption says the evaluation is on "two indoor scenes," but the table reports averages over two datasets comprising ten scenes; please rephrase.
  3. [§4.1, Table 5] The texture map size is given as 3×2048×2048 in §4.1 but as 2024×2024×3 in Table 5; please make the numbers consistent (2048×2048×3 appears to be intended).
  4. [Table 3] The row labels "Raw 0", "Raw 1", "Pruned 1", "Pruned T", "Pruned Sigmoid" are unclear; please define what each entry means (for example, whether M_T is constant 0, constant 1, or the sigmoid transmittance mask).
  5. [Eqs. (3), (6), §3.4] The symbol λ is used both for the D-SSIM weight in Eq. (3) and for the texture-loss weight in Eq. (6), and the sentence about setting λ to zero after densification is ambiguous; please use distinct symbols or clarify which loss term is being disabled.
  6. [Supplementary A] The method introduces several hand-set thresholds (scale threshold 0.01×D, K=500K, α_normal=20, α_area=50, α_group=100, warm-up 3k, k=20, adjacent-angle 45°), but only λ is ablated; a sensitivity analysis or at least a discussion of stability across scenes would strengthen the paper.
  7. [General] The manuscript does not state whether code, trained models, or the web viewer will be released; please include an availability statement for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the hybrid representation's Gaussian-count reduction is produced by joint optimization, not by construction or by self-citation.

full rationale

The claimed derivation chain is: extract a mesh with PGSR, prune it using normal-variation/angle/area metrics, bake a texture from training views, then jointly optimize Gaussians and the textured mesh under a photometric loss plus a transmittance-aware texture loss. None of the reported quantities—Gaussian count, PSNR, SSIM, LPIPS, or FPS—is defined in terms of an optimized parameter, and no fitted value is renamed as a prediction. The reduced Gaussian count follows from the objective that lets the textured mesh dominate pixels with high transmittance; this is the intended optimization mechanism, not an equality that assumes the conclusion. Self-citations to VastGaussian and Decoupling Appearance Variations appear only in a related-work survey sentence and are not load-bearing for the central claim. The FPS measurements are limited to the CUDA rasterizer routine and exclude mesh rasterization and compositing; that is a genuine evaluation-protocol weakness that could affect the efficiency comparison, but it is not circularity in the derivation. No passage was found in which Eq. X equals Eq. Y by construction, nor any self-citation chain that forces the result. The method is benchmarked against external datasets with external baselines, so the result is not circular.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on hand-set hyperparameters (pruning thresholds, loss weight, warm-up length, mask sharpness) and on the assumption that PGSR mesh geometry is reliable in retained flat regions. The efficiency gain is empirical and benchmarked, not derived from first principles.

free parameters (6)
  • Mesh pruning scale threshold = 0.01 x camera extent
    Gaussians larger than this are removed before TSDF extraction; hand-set in Supp. A and directly determines how much geometry is assigned to mesh.
  • QSlim target triangle count K = 500K
    Decimation target in Supp. A; controls mesh complexity before pruning.
  • Mesh pruning percentages (alpha_normal, alpha_area, alpha_group) = 20, 50, 100
    Percentiles of normal variation, triangle area, and connected-component size removed; hand-set defaults in Supp. A.
  • Adjacent angle threshold = 45 degrees
    Triangles with adjacent-face angles above this threshold are pruned to remove complex geometry and flat-region boundaries.
  • Transmittance mask sharpness k = 20
    Sigmoid steepness in Eq. 5; controls the soft threshold that decides mesh vs Gaussian dominance per pixel.
  • Texture loss weight and warm-up iterations = lambda=0.1, warm-up=3000
    Balance between L_c and L_t and the number of warm-up iterations; hand-set in Supp. A and ablated in Tab. 4.
assumptions (3)
  • domain assumption PGSR-extracted mesh geometry is accurate enough in retained flat regions after pruning
    Texture supervision bakes ground-truth colors onto the mesh; if retained geometry is wrong, the texture is distorted. Supp. B admits the pruned mesh still has geometric errors.
  • domain assumption Opaque mesh with fixed depth is a valid background model
    The composite rendering (Eq. I_h = I_gs + T*I_m) assumes the mesh renders as an opaque surface behind Gaussians; transparent or reflective surfaces are not modeled.
  • ad hoc to paper Transmittance threshold 0.5 reliably separates foreground objects from background mesh
    The sigmoid in Eq. 5 with k=20 makes a sharp binary decision; its validity is only demonstrated by the dataset experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Mesh-Gaussian Representation for Efficient Indoor Scene Reconstruction." pith.science (2026). https://pith.science/paper/C4SJ7UEF

@misc{pith2026250606988,
  author       = {Pith},
  title        = {Pith review of: Hybrid Mesh-Gaussian Representation for Efficient Indoor Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4SJ7UEF}},
  note         = {Machine review of arXiv:2506.06988}
}
read the original abstract

3D Gaussian splatting (3DGS) has demonstrated exceptional performance in image-based 3D reconstruction and real-time rendering. However, regions with complex textures require numerous Gaussians to capture significant color variations accurately, leading to inefficiencies in rendering speed. To address this challenge, we introduce a hybrid representation for indoor scenes that combines 3DGS with textured meshes. Our approach uses textured meshes to handle texture-rich flat areas, while retaining Gaussians to model intricate geometries. The proposed method begins by pruning and refining the extracted mesh to eliminate geometrically complex regions. We then employ a joint optimization for 3DGS and mesh, incorporating a warm-up strategy and transmittance-aware supervision to balance their contributions seamlessly.Extensive experiments demonstrate that the hybrid representation maintains comparable rendering quality and achieves superior frames per second FPS with fewer Gaussian primitives.

Figures

Figures reproduced from arXiv: 2506.06988 by the authors.

Figure 1
Figure 1. Visualization of Gaussian distributions and corresponding [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method pipeline. We first use normal, angle, and area size metrics to remove the meshes representing the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the rendered images and Gaussian distribution of various methods. Our hybrid representation effectively reduces [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of the rendered images of 3DGS and our method. The hybrid representation demonstrates sharper renderings compared [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Illustration of incorrect color projection and the final ren [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of mesh before and after pruning. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Rendered images and Gaussian distribution from five perspectives of Drjohnson. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Rendered images and Gaussian distribution from five perspectives of Creepyattic. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    Revising densification in gaussian splatting.CoRR, abs/2404.06109,

    [Bul`oet al., 2024 ] Samuel Rota Bul `o, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting.CoRR, abs/2404.06109,

    arXiv 2024

  2. [3]

    PGSR: planar-based gaussian splatting for efficient and high- fidelity surface reconstruction.CoRR, abs/2406.06521,

    [Chenet al., 2024a ] 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.CoRR, abs/2406.06521,

    arXiv

  3. [5]

    Headgas: Real-time animatable head avatars via 3d gaussian splatting

    [Dhamoet al., 2024 ] Helisa Dhamo, Yinyu Nie, Arthur Moreau, Jifei Song, Richard Shaw, Yiren Zhou, and Ed- uardo P ´erez-Pellitero. Headgas: Real-time animatable head avatars via 3d gaussian splatting. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sat- tler, and G ¨ul Varol, editors,Computer Vision - ECCV 2024 - 18th European Conf...

    work page 2024

  4. [6]

    Lightgaus- sian: Unbounded 3d gaussian compression with 15x re- duction and 200+ FPS.CoRR, abs/2311.17245,

    [Fanet al., 2023 ] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaus- sian: Unbounded 3d gaussian compression with 15x re- duction and 200+ FPS.CoRR, abs/2311.17245,

    arXiv 2023

  5. [7]

    Trim 3d gaussian splatting for accurate geometry representation.CoRR, abs/2406.07499,

    [Fanet al., 2024 ] Lue Fan, Yuxue Yang, Minxing Li, Hong- sheng Li, and Zhaoxiang Zhang. Trim 3d gaussian splatting for accurate geometry representation.CoRR, abs/2406.07499,

    arXiv 2024

  6. [8]

    Mini-splatting: Representing scenes with a constrained number of gaussians

    [Fang and Wang, 2024] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G ¨ul Varol, editors,Computer Vision - ECCV 2024 - 18th Eu- ropean Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXVI...

    work page 2024

  7. [9]

    Mesh- based gaussian splatting for real-time large-scale deforma- tion.CoRR, abs/2402.04796,

    [Gaoet al., 2024 ] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh- based gaussian splatting for real-time large-scale deforma- tion.CoRR, abs/2402.04796,

    arXiv 2024

  8. [10]

    Sugar: Surface-aligned gaussian splatting for ef- ficient 3d mesh reconstruction and high-quality mesh ren- dering

    [Gu´edon and Lepetit, 2024] Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for ef- ficient 3d mesh reconstruction and high-quality mesh ren- dering. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5354–5363. IEEE,

    work page 2024

Show all 38 references
  1. [11]

    Vmesh: Hybrid volume-mesh representation for efficient view syn- thesis

    [Guoet al., 2023 ] Yuan-Chen Guo, Yan-Pei Cao, Chen Wang, Yu He, Ying Shan, and Song-Hai Zhang. Vmesh: Hybrid volume-mesh representation for efficient view syn- thesis. In June Kim, Ming C. Lin, and Bernd Bickel, edi- tors,SIGGRAPH Asia 2023 Conference Papers, SA 2023, Sydney,...

    2023

  2. [12]

    [Hedmanet al., 2018 ] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel J. Brostow. Deep blending for free-viewpoint image-based rendering.ACM Trans. Graph., 37(6):257,

    2018

  3. [14]

    In- tegrating meshes and 3d gaussians for indoor scene reconstruction with SAM mask guidance.CoRR, abs/2407.16173,

    [Kim and Lim, 2024] Jiyeop Kim and Jongwoo Lim. In- tegrating meshes and 3d gaussians for indoor scene reconstruction with SAM mask guidance.CoRR, abs/2407.16173,

    2024 arXiv

  4. [15]

    Mod- ular primitives for high-performance differentiable render- ing.ACM Transactions on Graphics, 39(6),

    [Laineet al., 2020 ] Samuli Laine, Janne Hellsten, Tero Kar- ras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Mod- ular primitives for high-performance differentiable render- ing.ACM Transactions on Graphics, 39(6),

    2020

  5. [17]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

    [Linet al., 2024 ] 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 reconstruction. InIEEE/CVF Conference on Computer Vision and Pattern ...

    2024

  6. [18]

    Decoupling appearance variations with 3d consistent features in gaussian splatting

    [Linet al., 2025 ] Jiaqi Lin, Zhihao Li, Binxiao Huang, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Xiaofei Wu, Feng- long Song, and Wenming Yang. Decoupling appearance variations with 3d consistent features in gaussian splatting. arXiv preprint arXiv:2501.10788,

    2025 arXiv

  7. [20]

    Gaussianhair: Hair model- ing and rendering with light-aware gaussians.CoRR, abs/2402.10483,

    [Luoet al., 2024 ] Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair model- ing and rendering with light-aware gaussians.CoRR, abs/2402.10483,

    2024 arXiv

  8. [21]

    Taming 3dgs: High-quality radiance fields with limited resources

    [Mallicket al., 2024 ] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Francisco Vicente Carrasco, Markus Steinberger, and Fernando De la Torre. Taming 3dgs: High-quality radiance fields with limited resources. CoRR, abs/2406.15643,

    2024 arXiv

  9. [22]

    Srini- vasan, Matthew Tancik, Jonathan T

    [Mildenhallet al., 2020 ] Ben Mildenhall, Pratul P. Srini- vasan, Matthew Tancik, Jonathan T. Barron, Ravi Ra- mamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, e...

    2020

  10. [23]

    Reducing the memory footprint of 3d gaussian splatting.Proc

    [Papantonakiset al., 2024 ] Panagiotis Papantonakis, Geor- gios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting.Proc. ACM Comput. Graph. Interact. Tech., 7(1):16:1–16:17,

    2024

  11. [24]

    Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians

    [Qianet al., 2024 ] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, CVPR 2024, S...

    2024

  12. [25]

    Splattingavatar: Realistic real-time hu- man avatars with mesh-embedded gaussian splatting

    [Shaoet al., 2024 ] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. Splattingavatar: Realistic real-time hu- man avatars with mesh-embedded gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2024

  13. [26]

    [Straubet al., 2019 ] 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 Ca...

    2019 arXiv

  14. [28]

    Hybridnerf: Efficient neural rendering via adaptive volumetric surfaces

    [Turkiet al., 2024 ] Haithem Turki, Vasu Agrawal, Samuel Rota Bul `o, Lorenzo Porzi, Peter Kontschieder, Deva Ramanan, Michael Zollh ¨ofer, and Christian Richardt. Hybridnerf: Efficient neural rendering via adaptive volumetric surfaces. InIEEE/CVF Conference on Computer Vision...

    2024

  15. [29]

    Games: Mesh-based adapting and modification of gaussian splatting.CoRR, abs/2402.01459,

    [Waczynskaet al., 2024 ] Joanna Waczynska, Piotr Borycki, Slawomir Konrad Tadeja, Jacek Tabor, and Przemyslaw Spurek. Games: Mesh-based adapting and modification of gaussian splatting.CoRR, abs/2402.01459,

    2024 arXiv

  16. [30]

    Let there be color! — Large-scale texturing of 3D reconstructions

    [Waechteret al., 2014 ] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! — Large-scale texturing of 3D reconstructions. InProceedings of the Eu- ropean Conference on Computer Vision. Springer,

    2014

  17. [32]

    Implicit gaussian splatting with efficient multi-level tri- plane representation.CoRR, abs/2408.10041,

    [Wu and Tuytelaars, 2024] Minye Wu and Tinne Tuytelaars. Implicit gaussian splatting with efficient multi-level tri- plane representation.CoRR, abs/2408.10041,

    2024 arXiv

  18. [33]

    4d gaussian splatting for real-time dynamic scene rendering

    [Wuet al., 2024 ] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. InIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, CVPR 2024, Seattle...

    2024

  19. [34]

    Scannet++: A high-fidelity dataset of 3d indoor scenes

    [Yeshwanthet al., 2023 ] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InIEEE/CVF In- ternational Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 12–22. IEEE,

    2023

  20. [35]

    GSDF: 3dgs meets SDF for improved rendering and reconstruction.CoRR, abs/2403.16964,

    [Yuet al., 2024a ] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. GSDF: 3dgs meets SDF for improved rendering and reconstruction.CoRR, abs/2403.16964,

    arXiv

  21. [36]

    Mip-splatting: Alias- free 3d gaussian splatting

    [Yuet al., 2024b ] 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

  22. [37]

    Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting

    [Zhouet al., 2024 ] Shijie Zhou, Zhiwen Fan, Dejia Xu, Hao- ran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In Ales Leonardis, Elisa Ricci, St...

    2024

  23. [38]

    Algorithm 1:The overall algorithm. Input:ImagesI, MeshM Output:3DGSGS, MeshM, Texture mapT 1Stage 1: Mesh pruning 2Normal:M N =StableNormal(I, αnormal); 3Angle:M A =AdjacentAngle(M,45°); 4Area size:M S =AreaSize(M, αarea); 5Prune mask:M P =M N &&M A &&M S ; 6Pruned Mesh:M P =S...

    2018

  24. [2014]

    Schwing, and Shenlong Wang

    [Wenet al., 2024 ] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G. Schwing, and Shenlong Wang. Goma- vatar: Efficient animatable human modeling from monoc- ular video using gaussians-on-mesh. InIEEE/CVF Con- ference on Computer Vision and Pattern Recognition, CVPR 2024, ...

    2024

  25. [2018]

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

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

    2023

  26. [2019]

    HAHA: highly articulated gaussian human avatars with textured mesh prior.CoRR, abs/2404.01053,

    [Svitovet al., 2024 ] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. HAHA: highly articulated gaussian human avatars with textured mesh prior.CoRR, abs/2404.01053,

    2024 arXiv

  27. [2020]

    Dreamscene: 3d gaussian-based text- to-3d scene generation via formation pattern sampling

    [Liet al., 2024 ] Haoran Li, Haolin Shi, Wenli Zhang, Wen- jun Wu, Yong Liao, Lin Wang, Lik-Hang Lee, and Pengyuan Zhou. Dreamscene: 3d gaussian-based text- to-3d scene generation via formation pattern sampling. CoRR, abs/2404.03575,

    2024 arXiv

  28. [2023]

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

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

    arXiv

  29. [2024]

    Meshgs: Adaptive mesh-aligned gaussian splatting for high-quality rendering,

    [Choiet al., 2024 ] Jaehoon Choi, Yonghan Lee, Hyungtae Lee, Heesung Kwon, and Dinesh Manocha. Meshgs: Adaptive mesh-aligned gaussian splatting for high-quality rendering,

    2024

  30. [2025]

    Humangaussian: Text-driven 3d human gener- ation with gaussian splatting

    [Liuet al., 2024 ] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Zi- wei Liu. Humangaussian: Text-driven 3d human gener- ation with gaussian splatting. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattl...

    2024

Pith tools

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