Pith. sign in

REVIEW 5 major objections 7 minor 2 cited by

GDGS: 3D Gaussian Splatting Via Geometry-Guided Initialization And Dynamic Density Control

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read GDGS adds geometry-guided initialization, surface-aligned optimization, and region-aware density control to 3D Gaussian Splatting, claiming higher perceptual fidelity than the default pipeline while preserving real-time rendering.

desk verdict A plausible incremental improvement over 3DGS that cannot be evaluated as written because the paper never defines where the ground-truth centers or the surface mesh come from. read the letter →

arxiv 2507.00363 v1 pith:VD532ZMD submitted 2025-07-01 cs.CV

Xingjun Wang , Lianlei Shan This is my paper
classification cs.CV
keywords 3DGaussianSplattingnovelviewsynthesisgeometry-guidedinitializationsurfacealignmentadaptivedensitycontrolreal-timerenderingStructure-from-Motion
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 sets out to show that the standard 3D Gaussian Splatting pipeline leaves quality on the table in three places: how Gaussians are initialized, how they are guided toward a coherent surface, and how density is added or removed. Its proposed GDGS pipeline replaces raw SfM initialization with an MLP that predicts Gaussian centers, adds a surface-aligned optimization loss that pulls each Gaussian just outside the nearest mesh point with a normal-aligned orientation, and replaces global adaptive density control with a grid-based mechanism that clones Gaussians only in high-detail regions and spreads them out elsewhere. The authors report that at the same 30,000-iteration budget their method exceeds default 3DGS on average PSNR, SSIM, and LPIPS across Mip-NeRF360, Tanks & Temples, and Deep Blending, and that it shows gains already at 7,000 iterations. A careful reader cares because this would mean explicit splat primitives can reach high perceptual quality at real-time frame rates, without the per-ray cost of implicit neural fields.

What carries the argument

The load-bearing object is the training-time pipeline rather than any single new renderer. Its three mechanisms are: (1) a geometry-guided initialization MLP that ingests SfM points plus camera parameters and outputs Gaussian centers, trained to minimize $L_{init}$ against ground-truth centers $\mu^{gt}$; (2) a surface-aligned optimization stage where each Gaussian is pulled toward the nearest mesh point $\mu^{mesh}_i$ with a small outward offset $dN$ and its orientation is aligned to the normal $N$, using distance, alignment, and surface-fitting losses; and (3) a Dynamic Adaptive Density Control that divides the scene into fixed 3D grid regions $R_k$, identifies non-uniform regions by variance in Gaussian density, enriches them by cloning high-gradient Gaussians, and applies the Top-20 dispersion loss $L_{top20} = \frac{1}{|G_{top20}|}\sum_{x_i,x_j \in G_{top20}} \|x_i - x_j\|^2$ to decluster uniform regions. The paper's argument is that these mechanisms convert an unstructured Gaussian soup into an ordered, surface-aligned configuration with detail concentrated where it matters.

What would settle it

Run the released code on a Tanks & Temples scene and instrument Eqs. (2)--(7) to record where $\mu^{mesh}_i$, the normal $N$, and $\mu^{gt}_i$ come from; if no mesh file or ground-truth centers are loaded, then the surface-aligned and initialization losses are computed from the same SfM points used as input, and the reported quality gain is not attributable to the described external geometry.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that ill-posed initialization and spatially uniform density control are the two main bottlenecks in 3DGS, and both are addressable. The GDGS pipeline first trains an MLP to map the SfM point cloud and camera parameters to Gaussian centers, minimizing $L_{init} = \frac{1}{N}\sum_{i=1}^N \|\mu^{pred}_i - \mu^{gt}_i\|^2$. It then refines each Gaussian toward the target $\mu^{target}_i = \mu^{mesh}_i + d N$, the nearest surface mesh point shifted by a positive offset along the normal, using a composite loss $L = \lambda_d L_{dist} + \lambda_a L_{align} + \lambda_s L_{surface}$. Finally, Dynamic Adaptive Density Control tiles the scene into fixed regions $R_k$, clones Gaussians in high-gradient non-uniform regions, and applies the Top-20 dispersion loss $L_{top20}$ to decluster uniform ones. In the paper's reported numbers, the 30K variant reaches PSNR 28.36, SSIM 0.834, and LPIPS 0.191 on average over the three main benchmarks, beating the 3DGS-30K scores of 27.87, 0.827, and 0.213 respectively.

Load-bearing premise

The method relies on a surface mesh with normals and on 'ground truth' Gaussian centers as supervision, but the paper never says where they come from, so the described training procedure is under-specified.

Editorial extensions

If this is right

  • At equal training budgets, the rendered scenes would come out sharper, with fewer background artifacts, while still rasterizing in real time.
  • If the claim about 7K-iteration results holds, high-quality splatting becomes practical for settings where only a short training pass is affordable.
  • Region-aware density control implies that compute and memory are spent where detail actually is, which could make very large scenes tractable if regions scale sensibly.
  • Because Gaussians end up aligned to a surface with consistent normals, the trained representation is closer to a mesh, easing tasks like mesh extraction and surface editing.
  • The MLP initialization means the method no longer depends on the quality of the initial SfM point cloud as heavily, since the network learns to predict better centers.

Reading between the lines

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

  • Editorial inference: The paper never specifies the origin of $\mu^{gt}$ in Eq. (1) or the surface mesh in Eq. (2); a fair implementation may need to derive these from dense stereo, depth sensors, or the SfM points themselves. If the latter, the initialization stage is partly circular, and the reported gains would then be attributed to the density control and alignment losses rather than to the MLP
  • Editorial inference: Because the ablation shows removing the region-aware adjustments slightly improves SSIM and LPIPS while lowering PSNR, the Top-20 mechanism may be trading off structural metrics for perceptual ones; a testable variant would sweep the non-uniformity threshold $\delta_u$ to find a Pareto-optimal balance.
  • Editorial inference: The proposed region-based cloning and dispersion losses are generic and could be grafted onto other point-based renderers or even depth-supervised NeRF variants that output normals, offering a way to test the density-control mechanism in isolation.
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

5 major / 7 minor

Summary. The paper proposes GDGS, a modification of 3D Gaussian Splatting with three components: an MLP that predicts initial Gaussian centers from SfM points, a surface-aligned optimization that pushes Gaussians toward mesh surfaces along normals, and a dynamic adaptive density control (ADC) that clones or spreads Gaussians based on fixed grid regions and a 'Top 20 Loss'. Experiments on Mip-NeRF360, Tanks & Temples, Deep Blending, and Blender report improved PSNR, SSIM, and LPIPS over 3DGS-30K, and ablations are intended to justify each component.

Significance. The reported gains are potentially useful: for example, Table 1 shows ours-30K improves Tanks & Temples PSNR from 23.14 to 24.39 and LPIPS from 0.183 to 0.151 relative to 3DGS-30K, while retaining the real-time rasterization backbone. The paper evaluates on standard benchmarks and compares against common baselines, which is appropriate for the claim. However, the significance is currently conditional: two of the three core components depend on supervision signals (ground-truth Gaussian centers and surface meshes) that are never sourced, and the missing hyperparameters and missing full-model ablation row prevent independent verification.

major comments (5)
  1. [Section 3.2, Eq. (1)] The initialization loss Linit is defined against ground-truth Gaussian centers mu_gt, but the manuscript never states where these centers come from. The Mip-NeRF360, Tanks & Temples, and Deep Blending datasets provide images, camera poses, and SfM point clouds, not per-Gaussian ground-truth centers. If mu_gt are the same SfM points fed to the MLP, Eq. (1) becomes a near-identity regression that cannot explain the reported gains; if mu_gt are obtained from a pre-trained 3DGS model or an external reconstruction, that procedure is not described. The first contribution is therefore not reproducible as written.
  2. [Section 3.2, Eqs. (2), (5)-(7)] Eq. (2) moves each Gaussian toward mu_mesh, the nearest point on a surface mesh, plus an offset along the normal N, and Eqs. (5)-(7) all depend on this mesh. No source for a surface mesh is provided, no mesh-reconstruction step is described in Section 4.1, and the benchmark scenes do not ship meshes. Consequently Ldist, Lalign, and Lsurface are undefined for the reported experiments, and the surface-aligned optimization stage cannot be instantiated.
  3. [Section 3.3, Eqs. (8)-(12)] The dynamic ADC component is under-specified: the grid resolution and region count N, the thresholds delta_d, delta_g, delta_s, and delta_u, the loss weight lambda_top20, the offset d, and the definition of L_recon(R_k) are all absent. The 'density ratio of the Top 20 densest to Bottom 20 sparsest sectors' is also never formalized. Without these quantities the adaptive density control cannot be reproduced or checked.
  4. [Table 2, Section 4.3] The ablation table has no row for the full method, so the reader cannot compare NoInit, NoGloss, and NoDynADC against the complete model. The table also does not state which losses or components are removed in each variant. This weakens the paper's causal attribution of the Table 1 improvements to the three proposed innovations.
  5. [Section 3.2, Eq. (6)] The alignment loss uses a Gaussian 'orientation vector' U_i, but 3DGS represents Gaussian orientation with a covariance matrix (parameterized by a quaternion); the mapping from covariance to U_i is never defined, so Lalign is not computable as written.
minor comments (7)
  1. [Section 4.1] The sentence 'Training spans 30,000 iterations divided into three phases' appears twice; please remove the duplicate.
  2. [Section 3.3, Eq. (8)] The symbol delta_d is used for both the opacity threshold and the distance threshold; please use distinct symbols for these two quantities.
  3. [Section 3.3, Eq. (11)] Ltop20 sums over pairs xi, xj but normalizes by |G_top20| rather than |G_top20|^2; please clarify whether the pairwise sum is intentional and how it is normalized.
  4. [Tables 1 and 2] No error bars, confidence intervals, or multiple-seed results are reported; please state whether each number is a single run.
  5. [Section 4.2] The paper claims real-time rendering, but no frame-rate or timing measurements are reported anywhere in Section 4.
  6. [Section 3.2] The symbol N is used both for the number of Gaussians and for the surface normal; please resolve the notation conflict.
  7. [Title and Section 2] There are spacing typos in the title ('Spla tting') and in the Section 2 heading ('RELATED WORKS'); please correct them.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity; the central claim is an external benchmark comparison, not a derivation from its own assumptions.

full rationale

The paper's main claim is empirical: it reports higher SSIM and lower LPIPS than 3DGS-30K on Mip-NeRF360, Tanks & Temples, and Deep Blending (Section 4.2, Table 1). This is evaluated against external datasets and standard metrics, so the central result is not forced by the method's definitions. The geometry-guided initialization (Eq. 1) is trained against 'ground truth' centers mu_gt whose source is never specified, and the surface-aligned losses (Eqs. 2-7) require a surface mesh and normals that the benchmarks do not provide. However, the paper never states that mu_gt are the same SfM points fed to the MLP, nor does it define the mesh as a byproduct of its own output; underspecification is a reproducibility gap, not a circular reduction. The self-citations [23]-[53] appear only in a sentence listing segmentation works and are not load-bearing for any claimed result. The surface-aligned component is related to prior external work SuGaR [21], not to the authors' own unverified claims. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' prior work. Therefore no circular step can be exhibited with the required specificity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities are proposed. The central claim rests on several unstated supervision signals and hand-set thresholds; these choices are not derived from external benchmarks, so they weaken the contribution. The most serious items are the undefined ground-truth centers in Eq. (1), the unsourced mesh in Eq. (2), and the threshold parameters in Eqs. (8) through (12).

free parameters (5)
  • loss weights lambda_d, lambda_a, lambda_s = not specified
    Equation (4) balances distance, alignment, and surface losses; no values or tuning procedure are reported.
  • surface offset d = not specified
    Determines how far each Gaussian sits outside the mesh surface in Eq. (2); value is never given.
  • ADC thresholds delta_d, delta_g, delta_s, delta_u = not specified
    Equations (8), (9), and the region variance classification in Section 3.3 depend on thresholds that are never specified.
  • Top-20 loss weight lambda_top20 = not specified
    Equation (12) requires a weight balancing reconstruction and dispersion terms; the value is not given.
  • region count N and grid resolution = not specified
    Section 3.3 divides the scene into fixed regions R_k but never states the number of regions or the cell size.
assumptions (4)
  • ad hoc to paper A per-scene surface mesh with accurate normals is available whenever the method is applied.
    Eq. (2) and Eq. (7) rely on nearest mesh points and surface normals, but Section 3.2 never says where the mesh comes from.
  • ad hoc to paper Ground-truth Gaussian centers mu_gt exist and are available as training targets in Eq. (1).
    The initialization MLP is supervised by L_init in Eq. (1), but the paper never defines the ground-truth centers or how they are computed.
  • ad hoc to paper The density ratio of the Top 20 densest to Bottom 20 sparsest sectors and the variance threshold delta_u are well-defined and computable from fixed grid regions.
    Section 3.3 introduces Top 20 Loss and non-uniformity classification without an operational definition of the ratios, region sizes, or thresholds.
  • domain assumption The quantitative numbers attributed to prior methods in Table 1 are accurate and measured under the same protocol.
    The comparison relies on reported PSNR, SSIM, and LPIPS from previous papers; any inconsistency in protocols would change the stated improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GDGS: 3D Gaussian Splatting Via Geometry-Guided Initialization And Dynamic Density Control." pith.science (2026). https://pith.science/paper/VD532ZMD

@misc{pith2026250700363,
  author       = {Pith},
  title        = {Pith review of: GDGS: 3D Gaussian Splatting Via Geometry-Guided Initialization And Dynamic Density Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VD532ZMD}},
  note         = {Machine review of arXiv:2507.00363}
}
read the original abstract

We propose a method to enhance 3D Gaussian Splatting (3DGS)~\cite{Kerbl2023}, addressing challenges in initialization, optimization, and density control. Gaussian Splatting is an alternative for rendering realistic images while supporting real-time performance, and it has gained popularity due to its explicit 3D Gaussian representation. However, 3DGS heavily depends on accurate initialization and faces difficulties in optimizing unstructured Gaussian distributions into ordered surfaces, with limited adaptive density control mechanism proposed so far. Our first key contribution is a geometry-guided initialization to predict Gaussian parameters, ensuring precise placement and faster convergence. We then introduce a surface-aligned optimization strategy to refine Gaussian placement, improving geometric accuracy and aligning with the surface normals of the scene. Finally, we present a dynamic adaptive density control mechanism that adjusts Gaussian density based on regional complexity, for visual fidelity. These innovations enable our method to achieve high-fidelity real-time rendering and significant improvements in visual quality, even in complex scenes. Our method demonstrates comparable or superior results to state-of-the-art methods, rendering high-fidelity images in real time.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Role of Initialization in 3D Gaussian Splatting

    cs.CV 2026-03 unverdicted novelty 6.0 of 10

    Dense initialization of 3DGS does not consistently beat sparse SfM initialization for standard novel views, but improves off-trajectory generalization; no densification method wins everywhere.

  2. MAViE: A Multi-scale Adaptive Vision Encoder for Fine-grained Visual Perception and Efficient Multimodal Reasoning

    cs.CV 2026-07 reject novelty 4.0 of 10

    A multi-scale gated fusion plus question-conditioned token router is specified to cut VLM visual tokens ~80% while improving accuracy, but all reported gains are simulated placeholders.

Reference graph

Works this paper leans on

61 extracted references · 56 canonical work pages · cited by 2 Pith papers

  1. [1]

    3D Gaussian Splatting (3DGS) [1]has emerged as a cutting-edge approach for capturing and render- ing 3D scenes from novel perspectives

    INTRODUCTION Novel view synthesis is a fundamental task in computer vi- sion and graphics. 3D Gaussian Splatting (3DGS) [1]has emerged as a cutting-edge approach for capturing and render- ing 3D scenes from novel perspectives. Unlike NeRFs [2], which rely on MLPs which are computationally intensive and resource-demanding, 3DGS directly models scenes using...

    work page

  2. [2]

    We discuss related works in traditional scene reconstruction, neural rendering, and point-based rendering

    RELA TED WORKS Our research builds on 3D Gaussian Splatting (3DGS) [1]. We discuss related works in traditional scene reconstruction, neural rendering, and point-based rendering. Early scene reconstruction methods leveraged light fields for novel-view synthesis [3, 4], progressing to unstructured captures [5]. Structure-from-Motion (SfM) [6]introduced spa...

    arXiv 2025

  3. [3]

    METHOD 3.1. Overview Our research builds upon the 3D Gaussian Splatting (3DGS) framework [1] by introducing additional steps to enhance its initialization, optimization, and adaptive density control as shown in Figure 1. Starting with input images and camera pa- rameters calibrated using Structure-from-Motion (SfM) [6], we use the resulting sparse point c...

    work page

  4. [4]

    IMPLEMENTA TION, RESULTS, AND EV ALUA TION Fig. 4. Qualitative comparison of reconstructed scenes for Ground Truth, ours-30K, and 3DGS-30K. The insets high- light fine details in both indoor and outdoor scenes, showing the superior reconstruction fidelity of our approach in preserv- ing structural details 4.1. Implementation All models are optimized on a ...

    work page

  5. [5]

    By using SfM data and an MLP for initialization, we achieve improved convergence and accuracy over the original 3DGS

    CONCLUSIONS Our enhanced 3D Gaussian Splatting method addresses lim- itations in initialization, optimization, and density control, maintaining real-time performance. By using SfM data and an MLP for initialization, we achieve improved convergence and accuracy over the original 3DGS. Our approach refines Gaussian placement for higher fidelity reconstructi...

    work page

  6. [6]

    3d gaussian splatting for real-time radiance field rendering,

    Bernhard Kerbl, Pavel Korshunov, Alexander Keller, Florian Bernard, Jan Eric Lenssen, and Carsten Dachs- bacher, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics (TOG), vol. 42, no. 4, pp. 1–16, Jul. 2023

    work page 2023

  7. [7]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    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,”European Conference on Computer Vi- sion (ECCV), vol. 12347, pp. 405–421, Aug. 2020

    work page 2020

  8. [8]

    The lumigraph,

    Steven J. Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F. Cohen, “The lumigraph,”ACM SIG- GRAPH Proceedings, vol. 30, pp. 43–54, Aug. 1996

    work page 1996

Show all 61 references
  1. [9]

    Light field rendering,

    Marc Levoy and Pat Hanrahan, “Light field rendering,” ACM SIGGRAPH Proceedings, vol. 30, pp. 31–42, Aug. 1996

    1996

  2. [10]

    Unstructured lu- migraph rendering,

    Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen, “Unstructured lu- migraph rendering,”ACM SIGGRAPH Proceedings, vol. 20, pp. 425–432, Aug. 2001

    2001

  3. [11]

    Photo tourism: Exploring photo collections in 3d,

    Noah Snavely, Steven M. Seitz, and Richard Szeliski, “Photo tourism: Exploring photo collections in 3d,” ACM SIGGRAPH Proceedings, vol. 25, pp. 835–846, Aug. 2006

    2006

  4. [12]

    Multi-view stereo for community photo collections,

    Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M. Seitz, “Multi-view stereo for community photo collections,”International Confer- ence on Computer Vision (ICCV), vol. 1, pp. 1–8, Oct. 2007

    2007

  5. [13]

    Depth synthesis and local warps for plausible image-based navigation,

    Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine- Hornung, and George Drettakis, “Depth synthesis and local warps for plausible image-based navigation,”ACM Transactions on Graphics (TOG), vol. 32, no. 3, pp. 1– 12, Aug. 2013

    2013

  6. [14]

    Floating tex- tures,

    Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent, “Floating tex- tures,”Computer Graphics Forum, vol. 27, no. 2, pp. 409–418, Jun. 2008

    2008

  7. [15]

    Deep blending for free-viewpoint image-based rendering,

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow, “Deep blending for free-viewpoint image-based rendering,” ACM Transactions on Graphics (TOG), vol. 37, no. 6, pp. 257:1–257:15, Dec. 2018

    2018

  8. [16]

    Point-based neural rendering with per-view optimization,

    Georgios Kopanas, Julien Philip, Thomas Leimk ¨uhler, and George Drettakis, “Point-based neural rendering with per-view optimization,”Computer Graphics Fo- rum, vol. 40, no. 4, pp. 29–43, Aug. 2021

    2021

  9. [17]

    Advances in neural rendering,

    Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul P. Srinivasan, Edgar Tretschk, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al., “Advances in neural rendering,”Computer Graphics Forum, vol. 41, no. 2, pp. 703–735, May. 2022

    2022

  10. [18]

    Point sample rendering,

    Jeff P. Grossman and William J. Dally, “Point sample rendering,”Rendering Techniques, vol. 1, pp. 181–192, Jun. 1998

    1998

  11. [19]

    High-quality surface splatting on today’s gpus,

    Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt, “High-quality surface splatting on today’s gpus,”Symposium on Point-Based Graphics (SPBG), vol. 10, pp. 17–24, Aug. 2005

    2005

  12. [20]

    Surfels: Surface elements as ren- dering primitives,

    Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross, “Surfels: Surface elements as ren- dering primitives,”ACM SIGGRAPH Proceedings, vol. 12, pp. 335–342, Jul. 2000

    2000

  13. [21]

    Surface splatting,

    Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross, “Surface splatting,”ACM SIG- GRAPH Proceedings, vol. 20, pp. 371–378, Jul. 2001

    2001

  14. [22]

    Neural point- based graphics,

    Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky, “Neural point- based graphics,”European Conference on Computer Vi- sion (ECCV), vol. 12363, pp. 696–712, Aug. 2020

    2020

  15. [23]

    Adop: Approximate differentiable one-pixel point ren- dering,

    Darius R ¨uckert, Linus Franke, and Marc Stamminger, “Adop: Approximate differentiable one-pixel point ren- dering,”ACM Transactions on Graphics (TOG), vol. 41, no. 4, pp. 99:1–99:14, Jul. 2022

    2022

  16. [24]

    Pulsar: Effi- cient sphere-based neural rendering,

    Christoph Lassner and Michael Zollh ¨ofer, “Pulsar: Effi- cient sphere-based neural rendering,”Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1440–1449, Jun. 2021

    2021

  17. [25]

    Instant neural graphics primitives with a multiresolution hash encoding,

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexander Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM Trans- actions on Graphics (TOG), vol. 41, no. 4, pp. 102:1– 102:15, Jul. 2022

    2022

  18. [26]

    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,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 5354–5363

    2024

  19. [27]

    Multi-scale 3d gaussian splatting for anti-aliased rendering,

    Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee, “Multi-scale 3d gaussian splatting for anti-aliased rendering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  20. [28]

    Densenet-based land cover classification network with deep fusion,

    Lianlei Shan and Weiqiang Wang, “Densenet-based land cover classification network with deep fusion,” IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1–5, 2021

    2021

  21. [29]

    Uhrsnet: A semantic segmentation network specifi- cally for ultra-high-resolution images,

    Lianlei Shan, Minglong Li, Xiaobin Li, Yang Bai, Ke Lv, Bin Luo, Si-Bao Chen, and Weiqiang Wang, “Uhrsnet: A semantic segmentation network specifi- cally for ultra-high-resolution images,” in2020 25th In- ternational Conference on Pattern Recognition (ICPR). IEEE, 2021, pp. 1460–1466

    2021

  22. [30]

    Class-incremental learning for semantic segmentation in aerial imagery via distillation in all aspects,

    Lianlei Shan, Weiqiang Wang, Ke Lv, and Bin Luo, “Class-incremental learning for semantic segmentation in aerial imagery via distillation in all aspects,”IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–12, 2021

    2021

  23. [31]

    De- couple the high-frequency and low-frequency informa- tion of images for semantic segmentation,

    Lianlei Shan, Xiaobin Li, and Weiqiang Wang, “De- couple the high-frequency and low-frequency informa- tion of images for semantic segmentation,” inICASSP 2021-2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 1805–1809

    2021

  24. [32]

    Mbnet: A multi- resolution branch network for semantic segmentation of ultra-high resolution images,

    Lianlei Shan and Weiqiang Wang, “Mbnet: A multi- resolution branch network for semantic segmentation of ultra-high resolution images,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 2589–2593

    2022

  25. [33]

    Class-incremental semantic segmentation of aerial im- ages via pixel-level feature generation and task-wise dis- tillation,

    Lianlei Shan, Weiqiang Wang, Ke Lv, and Bin Luo, “Class-incremental semantic segmentation of aerial im- ages via pixel-level feature generation and task-wise dis- tillation,”IEEE Transactions on Geoscience and Re- mote Sensing, vol. 60, pp. 1–17, 2022

    2022

  26. [34]

    Global-local attention network for se- mantic segmentation in aerial images,

    Minglong Li, Lianlei Shan, Xiaobin Li, Yang Bai, Dengji Zhou, Weiqiang Wang, Ke Lv, Bin Luo, and Si-Bao Chen, “Global-local attention network for se- mantic segmentation in aerial images,” in2020 25th In- ternational Conference on Pattern Recognition (ICPR). IEEE, 2021, pp. 5704–5711

    2021

  27. [35]

    Continual learn- ing for image segmentation with dynamic query,

    Weijia Wu, Yuzhong Zhao, Zhuang Li, Lianlei Shan, Hong Zhou, and Mike Zheng Shou, “Continual learn- ing for image segmentation with dynamic query,”IEEE Transactions on Circuits and Systems for Video Technol- ogy, vol. 34, no. 6, pp. 4874–4886, 2023

    2023

  28. [36]

    In- cremental few shot semantic segmentation via class- agnostic mask proposal and language-driven classifier,

    Leo Shan, Wenzhang Zhou, and Grace Zhao, “In- cremental few shot semantic segmentation via class- agnostic mask proposal and language-driven classifier,” inProceedings of the 31st ACM International Confer- ence on Multimedia, 2023, pp. 8561–8570

    2023

  29. [37]

    Boosting gen- eral trimap-free matting in the real-world image,

    Leo Shan Wenzhang Zhou Grace Zhao, “Boosting gen- eral trimap-free matting in the real-world image,”arXiv preprint arXiv:2405.17916, 2024

    2024 arXiv

  30. [38]

    A data-related patch pro- posal for semantic segmentation of aerial images,

    Lianlei Shan, Guiqin Zhao, Jun Xie, Peirui Cheng, Xi- aobin Li, and Zhepeng Wang, “A data-related patch pro- posal for semantic segmentation of aerial images,”IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1– 5, 2023

    2023

  31. [39]

    End- to-end remote sensing change detection of unregis- tered bi-temporal images for natural disasters,

    Guiqin Zhao, Lianlei Shan, and Weiqiang Wang, “End- to-end remote sensing change detection of unregis- tered bi-temporal images for natural disasters,” inIn- ternational Conference on Artificial Neural Networks. Springer, 2023, pp. 259–270

    2023

  32. [40]

    Lifelong learning and selective forgetting via contrastive strategy,

    Lianlei Shan, Wenzhang Zhou, Wei Li, and Xingyu Ding, “Lifelong learning and selective forgetting via contrastive strategy,”arXiv preprint arXiv:2405.18663, 2024

    2024 arXiv

  33. [41]

    Cognitive memory in large language mod- els,

    Lianlei Shan, Shixian Luo, Zezhou Zhu, Yu Yuan, and Yong Wu, “Cognitive memory in large language mod- els,”arXiv preprint arXiv:2504.02441, 2025

    2025 arXiv

  34. [42]

    Dlnet: A dual-level network with self-and cross-attention for high-resolution remote sensing seg- mentation,

    Weijun Meng, Lianlei Shan, Sugang Ma, Dan Liu, and Bin Hu, “Dlnet: A dual-level network with self-and cross-attention for high-resolution remote sensing seg- mentation,”Remote Sensing, vol. 17, no. 7, pp. 1119, 2025

    2025

  35. [43]

    Transform dual- branch attention net: Efficient semantic segmentation of ultra-high-resolution remote sensing images,

    Bingyun Du, Lianlei Shan, Xiaoyu Shao, Dongyou Zhang, Xinrui Wang, and Jiaxi Wu, “Transform dual- branch attention net: Efficient semantic segmentation of ultra-high-resolution remote sensing images,”Remote Sensing, vol. 17, no. 3, pp. 540, 2025

    2025

  36. [44]

    Fusing multitask models by recursive least squares,

    Xiaobin Li, Lianlei Shan, and Weiqiang Wang, “Fusing multitask models by recursive least squares,” inICASSP 2021-2021 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 3640–3644

    2021

  37. [45]

    Ldnet: Semantic seg- mentation of high-resolution images via learnable patch proposal and dynamic refinement,

    Yuyang Ji and Lianlei Shan, “Ldnet: Semantic seg- mentation of high-resolution images via learnable patch proposal and dynamic refinement,” in2024 IEEE Inter- national Conference on Multimedia and Expo (ICME). IEEE, 2024, pp. 1–6

    2024

  38. [46]

    Organizing background to explore latent classes for incremental few-shot semantic segmentation,

    Lianlei Shan, Wenzhang Zhou, Wei Li, and Xingyu Ding, “Organizing background to explore latent classes for incremental few-shot semantic segmentation,”arXiv preprint arXiv:2405.19568, 2024

    2024 arXiv

  39. [47]

    Edge-guided and class-balanced active learning for se- mantic segmentation of aerial images,

    Lianlei Shan, Weiqiang Wang, Ke Lv, and Bin Luo, “Edge-guided and class-balanced active learning for se- mantic segmentation of aerial images,”arXiv preprint arXiv:2405.18078, 2024

    2024 arXiv

  40. [48]

    The binary quan- tized neural network for dense prediction via specially designed upsampling and attention,

    Xingyu Ding, Lianlei Shan, Guiqin Zhao, Meiqi Wu, Wenzhang Zhou, and Wei Li, “The binary quan- tized neural network for dense prediction via specially designed upsampling and attention,”arXiv preprint arXiv:2405.17776, 2024

    2024 arXiv

  41. [49]

    Energy minimum regularization in continual learning,

    Xiaobin Li, Lianlei Shan, Minglong Li, and Weiqiang Wang, “Energy minimum regularization in continual learning,” in2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021, pp. 6404– 6409

    2021

  42. [50]

    Boosting semantic segmentation of aerial images via decoupled and multilevel compaction and dispersion,

    Lianlei Shan, Weiqiang Wang, Ke Lv, and Bin Luo, “Boosting semantic segmentation of aerial images via decoupled and multilevel compaction and dispersion,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–16, 2023

    2023

  43. [51]

    Synthetic lung x-ray generation through cross-attention and affinity transfor- mation,

    Ruochen Pi and Lianlei Shan, “Synthetic lung x-ray generation through cross-attention and affinity transfor- mation,”arXiv preprint arXiv:2503.07209, 2025

    2025 arXiv

  44. [52]

    Dynrsl- vlm: Enhancing autonomous driving perception with dynamic resolution vision-language models,

    Xirui Zhou, Lianlei Shan, and Xiaolin Gui, “Dynrsl- vlm: Enhancing autonomous driving perception with dynamic resolution vision-language models,”arXiv preprint arXiv:2503.11265, 2025

    2025 arXiv

  45. [53]

    Flexdataset: Crafting an- notated dataset generation for diverse applications,

    Ellen Yi-Ge and Leo Shawn, “Flexdataset: Crafting an- notated dataset generation for diverse applications,” in Proceedings of the AAAI Conference on Artificial Intel- ligence, 2025, vol. 39, pp. 9481–9489

    2025

  46. [54]

    Gmm-based comprehen- sive feature extraction and relative distance preserva- tion for few-shot cross-modal retrieval,

    Chengsong Sun, Weiping Li, Xiang Li, Yuankun Liu, and Lianlei Shan, “Gmm-based comprehen- sive feature extraction and relative distance preserva- tion for few-shot cross-modal retrieval,”arXiv preprint arXiv:2505.13306, 2025

    2025 arXiv

  47. [55]

    Llm-cot enhanced graph neural rec- ommendation with harmonized group policy optimiza- tion,

    Hailong Luo, Bin Wu, Hongyong Jia, Qingqing Zhu, and Lianlei Shan, “Llm-cot enhanced graph neural rec- ommendation with harmonized group policy optimiza- tion,”arXiv preprint arXiv:2505.12396, 2025

    2025

  48. [56]

    Geogrambench: Bench- marking the geometric program reasoning in modern llms,

    Shixian Luo, Zezhou Zhu, Yu Yuan, Yuncheng Yang, Lianlei Shan, and Yong Wu, “Geogrambench: Bench- marking the geometric program reasoning in modern llms,”arXiv preprint arXiv:2505.17653, 2025

    2025

  49. [57]

    Geolocsft: Efficient visual geolocation via supervised fine-tuning of multimodal foundation models,

    Qiang Yi and Lianlei Shan, “Geolocsft: Efficient visual geolocation via supervised fine-tuning of multimodal foundation models,”arXiv preprint arXiv:2506.01277, 2025

    2025 arXiv

  50. [58]

    F2net: A frequency- fused network for ultra-high resolution remote sensing segmentation,

    Hengzhi Chen, Liqian Feng, Wenhua Wu, Xiaogang Zhu, Shawn Leo, and Kun Hu, “F2net: A frequency- fused network for ultra-high resolution remote sensing segmentation,”arXiv preprint arXiv:2506.07847, 2025

    2025 arXiv

  51. [59]

    Plenoxels: Radiance fields without neural networks,

    David B. Fridovich-Keil, Matthew Tancik, Qin- hong Chen, Benjamin Recht, and Angjoo Kanazawa, “Plenoxels: Radiance fields without neural networks,” CVPR Proceedings, vol. 41, pp. 5491–5500, Jun. 2022

    2022

  52. [60]

    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,”CVPR Proceedings, pp. 5670–5679, Jun. 2022

    2022

  53. [61]

    Tanks and temples: Benchmarking large-scale scene reconstruction,

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun, “Tanks and temples: Benchmarking large-scale scene reconstruction,”ACM Transactions on Graphics (TOG), vol. 36, no. 4, pp. 1–13, Jul. 2017

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.