Pith. sign in

REVIEW 4 major objections 6 minor 51 references

AG$^2$aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editing

T0 review · 4 major / 6 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 This paper claims that arranging semantic features on a voxel-anchor graph, rather than on free Gaussians, yields cleaner and more accurate instance-level selection for 3D scene understanding and editing.

desk verdict Competent incremental system for instance-level selection in 3DGS; the anchor-graph propagation is a real contribution, but the causal claim needs threshold-sensitivity evidence. read the letter →

arxiv 2508.01740 v1 pith:DZ6TXN6P submitted 2025-08-03 cs.CV

Zhaonan Wang , Manyi Li , Changhe Tu This is my paper
classification cs.CV
keywords 3DGaussianSplattinganchorgraphsemanticfeaturesopen-vocabularysegmentationinstance-levelselectionsceneeditingfeaturepropagationunderstanding
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 tries to establish that how semantic features are organized in 3D matters as much as which features are distilled. Its proposal, AG2aussian, replaces the common practice of attaching language and SAM-distilled features to free-floating 3D Gaussians with an anchor graph: a sparse set of voxel-centered anchors, each tied to a handful of Gaussians and connected to spatial neighbors, with features smoothed by graph propagation. The claim is that this structure produces smooth, instance-aware feature fields and therefore clean, complete Gaussian selections when a user clicks on or types an object. That matters because downstream operations, such as removing an object, recoloring it, or simulating physics on it, fail when selection leaves stray Gaussians behind or drops interior ones. On the LERF-OVS benchmark the paper reports click mIoU rising from 73.46 (OpenGaussian) to 88.33 and text mIoU from 44.17 to 54.35, with consistent gains on Mip-NeRF360 and LLFF.

What carries the argument

The anchor graph (Sec. 3.2). Anchors sit at the centers of occupied cells of a three-level voxelization built from the SfM point cloud; each anchor holds a semantic feature $f \in \mathbb{R}^3$, a voxel size $l$, and $k = 5$ child Gaussians whose means and scales are pinned inside the anchor's voxel by $\mu = x + o\cdot l$ and $s = \mathrm{Sigmoid}(\hat{s})\cdot l$. The graph is formed by intra-voxel edges among all anchors inside the same top-layer voxel and inter-voxel edges to anchors in neighboring top-layer voxels. Feature refinement is carried by the Gaussian-kernel-weighted graph Laplacian loss $L_{\mathrm{prop}} = 2\,\mathrm{Tr}(F^\top L F) = \sum_{i,j} w_{ij}\|F_i - F_j\|^2$ with $w_{ij} = \exp(-\|F_i - F_j\|^2 / 2\tau^2)$ and $\tau = 0.05$, which smooths features along object surfaces while leaving large feature differences intact as boundaries. The same graph drives selection: queries grow a connected region from a seed anchor under an edge-weight stopping rule, and Union-Find clustering plus IoU-feature matching attaches CLIP language features to clusters. One structure thus does the work of regularization, propagation, selection, and editing localization.

What would settle it

Render the semantic feature map and run the click-growing rule with edge weight above 0.90 on a pair of adjacent, interpenetrating objects of similar appearance in an LERF-style scene: if the query region leaks across the touching surface, the spatial-prior assumption fails exactly where it is weakest. A cheaper check is to recompute LERF-OVS text-query mIoU with the propagation temperature $\tau$ varied over, say, 0.01, 0.1, and 0.2 and the click threshold over 0.8 and 0.95; if the gains collapse under small threshold shifts, the result is tuned to the default rather than carried by the graph structure.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that a semantic 3D Gaussian field organized as an anchor graph behaves differently from the same field attached to free Gaussians: it stops Gaussians from sprawling across object boundaries, disambiguates which Gaussian belongs to which instance, and lets a graph-Laplacian smoothing step sharpen within-object features while keeping boundary features distinct. The paper argues that this structure, not the feature distillation alone, causes its quantitative gains, because removing the graph operations drops text-query mIoU on LERF from 54.35 to 42.72, and replacing the anchor localization with a learnable codebook or removing the local constraint drops it to 40.51 or 30.45. It further reports that the structured representation is cheaper as well as cleaner: about 7.56 GB peak memory versus 16.81 GB for OpenGaussian, and roughly 515 FPS rendering versus 96 FPS.

Load-bearing premise

The load-bearing premise is that anchors inside the same top-layer voxel, or in neighboring voxels, belong to the same object, so Gaussian-kernel smoothing with $\tau = 0.05$ sharpens within-object features while leaving boundaries intact; when instances touch or interpenetrate, spatial edges cross object boundaries and the hand-set stopping thresholds are the only barrier to leakage.

Editorial extensions

If this is right

  • Click-based instance selection on LERF-OVS reaches a mean mIoU of 88.33 and boundary IoU of 73.58, and text-driven selection reaches 54.35 mIoU, roughly ten points above the strongest baseline on both queries.
  • Because interior Gaussians are selected along with surface ones, object removal leaves a smaller, more localizable artifact region, and the paper's graph-boundary localization combined with LaMa inpainting produces coherent edited scenes, including on Mip-NeRF360.
  • Physics simulation behaves correctly when the selected object is dragged: the stiff selected object detaches from the remaining scene, whereas baselines either leave it in place or drag along stray surrounding Gaussians.
  • The anchor-graph representation is computationally lighter than the free-Gaussian baseline, using about 7.56 GB peak memory and rendering at roughly 515 FPS versus 16.81 GB and 96 FPS for OpenGaussian.
  • Ablations attribute the gains to the structure itself: removing the anchor graph or the graph operations progressively lowers text-query mIoU from 54.35 to between 42.72 and 49.10.

Reading between the lines

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

  • If the anchor-graph organization is the active ingredient, a testable extension is porting the same voxel-anchor graph to other feature distillations, such as DINO or other open-set features, and checking whether the selection gains persist; the paper's ablations only vary the structure while keeping SAM and CLIP fixed.
  • The spatial prior implies a prediction the paper does not stress-test: instances that touch or interpenetrate, where intra-voxel edges cross object boundaries, should degrade more than separated instances, and a per-category IoU breakdown on cluttered scenes would quantify how much of the gain survives.
  • The fixed stopping rules (edge weight above 0.90 for clicks, $\epsilon - 0.1$ for text) suggest an unstated tunable: per-scene adaptive thresholds, or a learned stopping criterion, could extend the gains to scenes with noisier feature fields.
  • Because selection is expressed as a connected subgraph rather than a similarity threshold on point features, editing operations could be composed as graph operations, such as union, difference, or containment queries; the paper does not explore this.
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 AG^2aussian, a 3D Gaussian Splatting representation in which Gaussians are attached to anchors placed at multi-resolution voxel centers, and semantic features live on anchors rather than on free Gaussians. Features are refined by graph Laplacian propagation over voxel-adjacency edges (Sec. 3.2), followed by graph clustering and CLIP feature attachment (Sec. 3.3). The authors evaluate interactive click query, open-vocabulary text query, object removal, and physics simulation on LERF-OVS, Mip-NeRF360, and LLFF, reporting large mIoU/mBIoU gains over SAGA, GaussianGrouping, and OpenGaussian (e.g., LERF mean click mIoU 88.33 vs 73.46 for OpenGaussian; mean text mIoU 54.35 vs 44.17). Ablations separate the anchor structure, the propagation loss, and the region-growing query. The central claim is that organizing semantic features on the anchor-graph, rather than on free Gaussians, is the cause of the improved instance-level selection.

Significance. The anchor-graph idea is clear, the application coverage is broad, and the paper includes qualitative and quantitative comparisons together with a substantial supplementary. If the reported gains survive independent evaluation, this would be a useful step for structured semantic 3DGS: it reduces the number of Gaussians, provides a natural region-growing query mechanism, and supports editing and simulation with one representation. The ablation design is mostly sensible, and the supplementary gives per-scene tables and additional qualitative results. However, the paper currently lacks statistical reliability measures, threshold-sensitivity analysis, and an evaluation ground truth that is independent of the SAM masks used for supervision, so the magnitude of the claimed advantage and its causal attribution to the graph structure are not yet established at the level the paper asserts.

major comments (4)
  1. [Sec. 5.1 / Sec. 3.1 / Eq. (8)] The evaluation ground truth is the same source as the supervision signal. The GT masks are described as 'manually annotated with SAM'; the contrastive losses in Sec. 3.1 use SAM masks, and the matching score in Eq. (8) attaches language features by IoU matching to SAM masks. The reported mIoU/mBIoU therefore measure agreement with SAM masks, which are also the training target of the proposed method. This may inflate the absolute and relative numbers, especially for the text-query results. Please report results with human-verified or independently annotated masks, and state clearly whether all baselines also use SAM masks as their training target; if they do, the fair-comparison caveat should be stated prominently.
  2. [Sec. 4 / Eq. (7)] The method relies on several thresholds for which no validation protocol or sensitivity study is provided: the click region-growing edge threshold wij > 0.90, the text seed margin epsilon - 0.1, the propagation temperature tau = 0.05, and the anchor densification gradient threshold are all fixed. The causal claim that the anchor graph and propagation cause the gains is only supported by ablations that toggle Lprop and GraphSeg on/off (Tables 2 and 6), not by showing robustness to these thresholds. Please add sweeps over these parameters (e.g., tau in {0.01, 0.02, 0.05, 0.1, 0.2}, edge threshold in {0.7, 0.8, 0.9, 0.95}, and epsilon margin in {0.05, 0.1, 0.2}), and report the densification threshold used.
  3. [Tables 1, 4, 5] All quantitative comparisons are single runs with no variance estimates. Several per-scene numbers are close or even unfavorable to the proposed method (e.g., Table 1, kitchen click mBIoU: Ours 49.86 vs OpenGaussian 49.95), while the text claims 'consistent gains' across scenes. Without multiple seeds or a paired significance test, the conclusion that the anchor-graph structure reliably outperforms the baselines is not statistically supported. Please report mean +/- std over at least three training runs, or a paired test across scenes, and discuss per-scene exceptions.
  4. [Sec. 3.2] The graph construction connects anchors purely by spatial adjacency of top-layer voxels; when two instances touch or interpenetrate, intra- and inter-voxel edges cross object boundaries by construction. The only defenses are the feature-kernel weight, the region-growing threshold, and the text margin. The paper does not analyze this failure mode and reports no experiment specifically on tightly adjacent or interpenetrating objects. Please add an experiment (e.g., on the LERF kitchen/ramen scenes or a controlled pair of touching objects) reporting leakage rates and, if computable, the fraction of graph edges that connect different SAM instances.
minor comments (6)
  1. [Sec. 5.1] The phrase 'manually annotated with SAM' is ambiguous; clarify whether a human verified every mask or whether automatically generated SAM masks were used directly as ground truth.
  2. [Sec. 5.2] Please define how mIoU and mBIoU are aggregated (per-scene mean then average, or global average over all pixels) and state the number of click points and text prompts used per scene.
  3. [Tables 2 and 6] The row labels '#1', '#2', and '#3' are not self-contained; include the checkmark states in a legend so the reader can tell exactly which components are active in each row.
  4. [Sec. 3.2] The choice of tau = 0.05 deserves a sentence relating it to the typical L2 norm of the 3D semantic features; as written, a feature distance of about 0.07 already reduces the Gaussian weight to exp(-1), which makes long-range propagation extremely brittle.
  5. [Introduction / Code release] The string 'Code released in GitHub/AGGaussian' is not a usable link and appears to be a placeholder; provide an anonymized repository URL or state the release plan explicitly.
  6. [Sec. 5.3 / Fig. 5] The object-removal and artifact-inpainting results are qualitative only; a quantitative metric (e.g., LPIPS or mask-consistency of the inpainted region) would strengthen the editing claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the anchor-graph design is validated by component-wise ablations against baselines trained under the same supervision.

full rationale

No step in the paper's derivation chain is equivalent to its inputs by construction. The anchor-graph construction (Sec. 3.2) is an empirical spatial graph over voxelized anchors; the propagation loss Lprop (Eq. 7) is a standard Dirichlet-energy regularizer, and the query region-growing thresholds (0.90 and epsilon-0.1, Sec. 4) are inference-time hyperparameters, not fitted outputs renamed as predictions. The reported gains are supported by ablations (Tables 2 and 6) that isolate the graph propagation and graph segmentation from the anchor representation. The point that deserves note is that SAM masks are used both as training supervision (Sec. 3.1), as the source of attached CLIP features (Sec. 3.3), and as evaluation ground truth (Sec. 5.1). This is a supervised-evaluation overlap, not a by-construction equivalence: the method's binary instance maps are produced by learned-feature clustering and region growing, and the same SAM-mask supervision is given to the OpenGaussian, SAGA, and GsGrouping baselines, so the comparison remains informative. The paper's own stated limitations (glass/metal, SAM over-segmentation, Sec. 6) are empirical failure modes, not admissions of circularity. No self-citations are load-bearing, and no uniqueness theorem or ansatz is imported from the authors' prior work.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim inherits assumptions from Scaffold-GS (voxel-anchor regularization), OpenGaussian (contrastive distillation and codebook matching), and SAM/CLIP (mask and language quality). The model's own added assumption is that voxel-adjacency defines instance connectivity and that graph smoothing with a fixed kernel preserves boundaries. Hand-set thresholds (0.90, 0.1) and loss weights are the free parameters that the reported numbers rest on. No new physical entities or forces are introduced; anchors are algorithmic structures.

free parameters (6)
  • Propagation kernel temperature tau = 0.05
    Sets width of the Gaussian weighting used in the Laplacian propagation loss for all anchor features; chosen by hand, not validated on a held-out set.
  • Region-growing edge threshold (click query) = 0.90
    During click query, an anchor joins the selected region only if edge weight > 0.90; the threshold controls selection cleanness vs completeness and is fixed without validation protocol.
  • Text query seed margin = epsilon - 0.1
    Seed anchors for text query are those clusters whose language-feature similarity is within 0.1 of the maximum score; this manual margin affects which instances are selected.
  • Loss weights (lambda_in, lambda_is, lambda_ic, lambda_d, lambda_prop) = 0.5, 2.5, 0.25, 50, 0.01
    Balance regularization, intra-mask smoothing, inter-mask contrast, depth distortion, and propagation terms; tuned on the LERF scenes used for evaluation.
  • Child Gaussians per anchor k = 5
    Every anchor owns exactly k=5 Gaussians, which bounds expressiveness and memory; not ablated across k.
  • Densification gradient threshold = not specified
    New anchors are created when a child Gaussian's training gradient exceeds a threshold, but the threshold value is never reported, so the anchor count is not reproducible.
assumptions (5)
  • domain assumption SAM instance masks are a valid ground truth for object-level segmentation.
    The paper uses SAM ViT-H to produce multi-view masks for both supervision (Sec. 3.1) and evaluation (Sec. 5.2), assuming these masks correspond to true instances.
  • domain assumption Anchors within the same top-layer voxel or neighboring voxels are likely to belong to the same object instance, so feature propagation should be restricted to these edges.
    Sec. 3.2 builds intra-voxel and inter-voxel edges without checking whether the voxel boundary aligns with instance boundaries; if two objects touch, the graph mixes their features.
  • domain assumption CLIP and SAM embeddings encode semantics well enough that contrastive distillation and language matching work.
    Secs. 3.1 and 3.3 rely on OpenGaussian-style contrastive training and CLIP features; the paper inherits these assumptions rather than proving them.
  • domain assumption The voxelized anchor structure from Scaffold-GS is a sensible inductive bias for Gaussian regularization.
    Eq. 2 constrains Gaussian mean and scale to the anchor's voxel, an assumption the paper imports from Scaffold-GS and validates only via the w/o localization ablation.
  • ad hoc to paper Gaussian-kernel graph Laplacian propagation with tau=0.05 smooths within-object features while preserving boundaries.
    The Dirichlet-energy loss (Eq. 7) with this kernel is proposed in Sec. 3.2 without analysis showing boundary preservation; its benefit is shown only empirically in Table 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AG$^2$aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editing." pith.science (2026). https://pith.science/paper/DZ6TXN6P

@misc{pith2026250801740,
  author       = {Pith},
  title        = {Pith review of: AG$^2$aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZ6TXN6P}},
  note         = {Machine review of arXiv:2508.01740}
}
abstract

3D Gaussian Splatting (3DGS) has witnessed exponential adoption across diverse applications, driving a critical need for semantic-aware 3D Gaussian representations to enable scene understanding and editing tasks. Existing approaches typically attach semantic features to a collection of free Gaussians and distill the features via differentiable rendering, leading to noisy segmentation and a messy selection of Gaussians. In this paper, we introduce AG$^2$aussian, a novel framework that leverages an anchor-graph structure to organize semantic features and regulate Gaussian primitives. Our anchor-graph structure not only promotes compact and instance-aware Gaussian distributions, but also facilitates graph-based propagation, achieving a clean and accurate instance-level Gaussian selection. Extensive validation across four applications, i.e. interactive click-based query, open-vocabulary text-driven query, object removal editing, and physics simulation, demonstrates the advantages of our approach and its benefits to various applications. The experiments and ablation studies further evaluate the effectiveness of the key designs of our approach.

Figures

Figures reproduced from arXiv: 2508.01740 by the authors.

Figure 1
Figure 1. We propose AG2 aussian, an anchor-graph structured Gaussian splatting for instance-level 3D scene understanding and editing tasks. Compared to existing works that attach semantic features to a collection of free Gaussians, we construct an anchor-graph structure to organize the semantic features and regulate the associated Gaussians, resulting in a smooth feature distribution and a clean and accurate instance-level G… view at source ↗
Figure 2
Figure 2. Our AG2 aussian consists of three stages. The first stage performs anchor-gaussian growing, where we initialize the anchors based on multi-resolution voxelization and optimize the anchor-Gaussian via a differentiable rasterizer. The second stage constructs the anchor￾graph structure with sparse inter-voxel edges and intra-voxel edges, and adopts a graph-based propagation to refine the semantic features. Finally, the… view at source ↗
Figure 3
Figure 3. The rendered semantic feature map and object query results of our approach and related methods. The presented GT images are [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Object removal results. After deleting the selected Gaus [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 4
Figure 4. Figure 4: Click-based instance-level Gaussian selection and the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Physical simulation by applying the external forces (red arrows) to drag the objects. All methods are simulated with identical [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visual results of the ablation study for anchor-graph [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Physical simulation by applying the external forces (red [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Open-vocabulary 3D object selection on the LERF dataset [ [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 11
Figure 11. Figure 11: More editing results on MipNeRF360 [1] using our graph-based artifact localization technique. Recoloring & Insertion Insertion & Recoloring +Recoloring [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: More object recoloring and insertion editing results on [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Open-vocabulary 3D object selection on the Mip-NeRF360 dataset [ [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Scribbles-based 3D object selection on the LLFF dataset [ [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Ablation study results. We separately validated the importance of our key design for segmentation task, the Anchor-Gaussian [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages

  1. [1]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 6, 1, 2, 3, 4

    work page 2022

  2. [2]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021. 1, 2

    work page 2021

  3. [3]

    Segment any 3d gaus- sians

    Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xi- aopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaus- sians. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1971–1979, 2025. 2, 6, 7, 1

    work page 1971

  4. [4]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 2024. 1, 2

    work page 2024

  5. [5]

    Gaussianeditor: Swift and control- lable 3d editing with gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xi- aofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control- lable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21476–21485, 2024. 1, 2

    work page 2024

  6. [6]

    Hac: Hash-grid assisted context for 3d gaussian splatting compression

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac: Hash-grid assisted context for 3d gaussian splatting compression. In European Conference on Computer Vision, 2024. 2, 3

    work page 2024

  7. [7]

    Tracking anything with decoupled video segmentation

    Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexan- der Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1316–1326, 2023. 1, 2

    work page 2023

  8. [8]

    High-quality surface recon- struction using gaussian surfels

    Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface recon- struction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 2

    work page 2024

Show all 51 references
  1. [9]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, Zhangyang Wang, et al. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. Advances in neural information processing systems, 37: 140138–140158, 2024. 2

    2024

  2. [10]

    Mesh-based gaussian splat- ting for real-time large-scale deformation

    Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based gaussian splat- ting for real-time large-scale deformation. arXiv preprint arXiv:2402.04796, 2024. 2

    2024 arXiv

  3. [11]

    Towards realistic example- based modeling via 3d gaussian stitching

    Xinyu Gao, Ziyi Yang, Bingchen Gong, Xiaoguang Han, Sipeng Yang, and Xiaogang Jin. Towards realistic example- based modeling via 3d gaussian stitching. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26597–26607, 2025. 2

    2025

  4. [12]

    Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning

    Qiao Gu, Ali Kuwajerwala, Sacha Morin, Krishna Murthy Jatavallabhula, Bipasha Sen, Aditya Agarwal, Corban Rivera, William Paul, Kirsty Ellis, Rama Chellappa, et al. Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning. In 2024 IEEE International Conferenc...

    2024

  5. [13]

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

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 2

    2024

  6. [14]

    A moving least squares material point method with displacement disconti- nuity and two-way rigid body coupling

    Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement disconti- nuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG), 37(4):150, 2018. 6, 1

    2018

  7. [15]

    2d gaussian splatting for geometrically accu- rate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accu- rate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1–11, 2024. 1, 2, 4

    2024

  8. [16]

    Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors

    Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. arXiv preprint arXiv:2406.01476,

    arXiv

  9. [17]

    Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces

    Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaox- iao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 53...

    2024

  10. [18]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5, 8

    2023

  11. [19]

    Lerf: Language em- bedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language em- bedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729– 19739, 2023. 6, 7, 2, 3, 4, 5

    2023

  12. [20]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023. 2...

    2023

  13. [21]

    Discene: Object decoupling and interaction 9 modeling for complex scene generation

    Xiao-Lei Li, Haodong Li, Hao-Xiang Chen, Tai-Jiang Mu, and Shi-Min Hu. Discene: Object decoupling and interaction 9 modeling for complex scene generation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024. 1

    2024

  14. [22]

    Su- pergseg: Open-vocabulary 3d segmentation with structured super-gaussians, 2024

    Siyun Liang, Sen Wang, Kunyi Li, Michael Niemeyer, Ste- fano Gasperini, Nassir Navab, and Federico Tombari. Su- pergseg: Open-vocabulary 3d segmentation with structured super-gaussians, 2024. 4

    2024

  15. [23]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38–55. Springer...

    2024

  16. [24]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 2, 3, 4

    2024

  17. [25]

    Ever: Exact volumet- ric ellipsoid rendering for real-time view synthesis

    Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T Barron, and Yinda Zhang. Ever: Exact volumet- ric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2

    arXiv

  18. [26]

    Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar

    Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6, 1, 4

    2019

  19. [27]

    Langsplat: 3d language gaussian splatting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051–20060, 2024. 1, 2

    2024

  20. [28]

    Goi: Find 3d gaussians of interest with an optimizable open-vocabulary semantic-space hyperplane

    Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liu- juan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an optimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages ...

    2024

  21. [29]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

    2021

  22. [30]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2

    2024 arXiv

  23. [31]

    Schwing†, and Oliver Wang†

    Zhongzheng Ren, Aseem Agarwala †, Bryan Russell †, Alexander G. Schwing†, and Oliver Wang†. Neural volumet- ric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. († alphabetic ordering). 1

    2022

  24. [32]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016. 3

    2016

  25. [33]

    Aligning and prompting everything all at once for uni- versal visual perception

    Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for uni- versal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2024

  26. [34]

    Language embedded 3d gaussians for open- vocabulary scene understanding

    Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao- Hua Guan. Language embedded 3d gaussians for open- vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333–5343, 2024. 1, 2

    2024

  27. [35]

    Resolution-robust large mask inpainting with fourier convolutions

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winte...

    2022

  28. [36]

    Efficiency of a good but not linear set union algorithm

    Robert Endre Tarjan. Efficiency of a good but not linear set union algorithm. Journal of the ACM (JACM), 22(2):215– 225, 1975. 5

    1975

  29. [37]

    Games: Mesh-based adapt- ing and modification of gaussian splatting

    Joanna Waczy ´nska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapt- ing and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2

    2024 arXiv

  30. [38]

    Superpoint gaus- sian splatting for real-time high-fidelity dynamic scene re- construction

    Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaus- sian splatting for real-time high-fidelity dynamic scene re- construction. arXiv preprint arXiv:2406.03697, 2024. 2

    2024 arXiv

  31. [39]

    Recent advances in 3d gaussian splatting

    Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan- Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613– 642, 2024. 2

    2024

  32. [40]

    Gaussian head & shoulders: High fidelity neural upper body avatars with an- chor gaussian guided texture warping

    Tianhao Walter Wu, Jing Yang, Zhilin Guo, Jingyi Wan, Fangcheng Zhong, and Cengiz Oztireli. Gaussian head & shoulders: High fidelity neural upper body avatars with an- chor gaussian guided texture warping. In The Thirteenth International Conference on Learning Representations,

  33. [41]

    Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding.Advances in Neural Information Processing Systems, 37:19114–19138,

    Yanmin Wu, Jiarui Meng, Haijie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, et al. Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding.Advances in Neural Information Processing Systems, 37:19114–19138,

  34. [42]

    Physgaussian: Physics-integrated 3d gaussians for generative dynamics

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389–4398, 2024. 1, 2, 6

    2024

  35. [43]

    Texture-gs: Disentangling the geometry and texture for 3d gaussian splatting editing

    Tian-Xing Xu, Wenbo Hu, Yu-Kun Lai, Ying Shan, and Song-Hai Zhang. Texture-gs: Disentangling the geometry and texture for 3d gaussian splatting editing. In European Conference on Computer Vision, pages 37–53. Springer,

  36. [44]

    Llm- grounder: Open-vocabulary 3d visual grounding with large 10 language model as an agent

    Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm- grounder: Open-vocabulary 3d visual grounding with large 10 language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), page...

    2024

  37. [45]

    Reflective gaussian splatting

    Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting. arXiv preprint arXiv:2412.19282, 2024. 2

    2024 arXiv

  38. [46]

    Gaussian grouping: Segment and edit anything in 3d scenes

    Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. 1, 2, 6, 7

    2024

  39. [47]

    Stag4d: Spatial-temporal anchored generative 4d gaussians

    Yifei Zeng, Yanqin Jiang, Siyu Zhu, Yuanxun Lu, Youtian Lin, Hao Zhu, Weiming Hu, Xun Cao, and Yao Yao. Stag4d: Spatial-temporal anchored generative 4d gaussians. In European Conference on Computer Vision, pages 163–

  40. [48]

    Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling

    Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. CoRR, 2024. 1

    2024

  41. [49]

    Feature 3dgs: Supercharg- ing 3d gaussian splatting to enable distilled feature fields

    Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Ze- hao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharg- ing 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Comp...

    2024

  42. [50]

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

    Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern...

    2024

  43. [51]

    Ewa splatting

    Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223–238, 2002. 3 11 AG2aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editin...

    2002

Pith tools

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