Object-Centric 2D Gaussian Splatting: Background Removal and Occlusion-Aware Pruning for Compact Object Models

Didier Stricker; Marcel Rogge

REVIEW 3 major objections 5 minor 1 cited by

Object-Centric 2D Gaussian Splatting: Background Removal and Occlusion-Aware Pruning for Compact Object Models

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A segmentation-mask background loss and occlusion-aware pruning let 2D Gaussian Splatting reconstruct only a target object, yielding models up to 96% smaller and 71% faster to train with competitive quality.

desk verdict Occlusion-aware pruning is a genuinely useful trick, but the Mip-NeRF360 evaluation shares its masks with the training loss, so the 'competitive quality' claim should be read with caution. read the letter →

arxiv 2501.08174 v2 pith:XHLSOBF6 submitted 2025-01-14 cs.CV

Marcel Rogge , Didier Stricker This is my paper · ORCID

classification cs.CV

keywords object-centricreconstruction2DGaussiansplattingbackgroundlossocclusion-awarepruningsegmentationmasksmeshextractionnovelviewsynthesissurface

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 Gaussian Splatting can be made object-centric without sacrificing much quality: by feeding segmentation masks into the optimization loop, a scene-level reconstructor is forced to model only a target object, and by pruning Gaussians that never contribute to any rendered view, the resulting model is far smaller and faster to train. The method is built on 2D Gaussian Splatting (a surface-oriented variant in which geometry is represented as flat Gaussian discs), adding a background loss on accumulated opacity and masking the photometric loss. On Mip-NeRF360 it reports reducing the mean number of Gaussians from 2,053,425 to 73,179 (about 96% smaller) and cutting training time from 31.44 to 9.15 minutes (about 71% faster), with masked PSNR dropping from 28.31 to 25.69. A curious reader would care because compact object-centric models are the missing piece for editing, physics simulation, and other downstream tasks on single objects.

What carries the argument

The load-bearing object is the background loss $L_b = \frac{1}{h \cdot w}\sum [A_i \cdot (1 - M_i)]$, which drives accumulated $\alpha$ to zero outside the object mask, paired with a masked photometric loss $L_c^M = L_c(I_i \cdot M_i, R_i \cdot M_i)$ so color fitting does not counteract the background removal. The second mechanism is occlusion-aware pruning: the renderer records which Gaussians actually take part in $\alpha$ blending, and the adaptive density control removes any Gaussian that never contributes. These sit on top of the 2DGS depth-distortion and normal-consistency losses that keep the 2D Gaussians glued to surfaces, which is what makes mesh extraction possible.

What would settle it

Run the full method on a scene with perfectly known ground-truth masks (e.g., from synthetic rendering) and compare masked PSNR to the 2DGS baseline; if the gap remains larger than about 0.5 dB, the quality drop is not explained by mask errors, which the paper's robustness argument requires.

Watch

Extended reading notes

Core claim

The central claim is that two inexpensive changes to 2D Gaussian Splatting turn it into an object-level reconstructor. The first is a background loss that penalizes the rendered accumulated alpha wherever the segmentation mask is zero, pushing background Gaussians to become transparent so the automatic density control deletes them; the second is an occlusion-aware pruning rule that deletes Gaussians which never participate in the alpha blending of any training view. In the full method on the Mip-NeRF360 dataset, the mean number of Gaussians falls from 2,053,425 to 73,179 (a 96% reduction), training time falls from 31.44 to 9.15 minutes (a 71% reduction), and masked PSNR goes from 28.31 to 25.69. The authors attribute part of that quality drop to errors in the SAM 2 segmentation masks rather than to the optimization itself.

Load-bearing premise

The method assumes the input segmentation masks are accurate and consistent across all views; if object pixels are systematically labeled as background, the background loss penalizes correct object Gaussians and the masked photometric loss prevents the error from being visible.

Editorial extensions

If this is right

Object-specific Gaussian models become practical to train on a single consumer GPU, cutting training time by about 71% on scene-scale data.
Exported models are roughly 96% smaller, which shrinks storage and memory footprint from tens of megabytes to a few megabytes for typical objects.
The pruning step gives a near-free reduction of about 10% of Gaussians even when no masks are used, with no measurable quality change.
Because the object is already isolated, meshes and Gaussians can be plugged directly into appearance editing and physics simulation without extra segmentation or culling.
The pruning rule works in the original 3DGS as well, indicating the visibility-tracking idea transfers to Gaussian methods beyond 2DGS.

Reading between the lines

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

The visibility-tracking pruning is a general inference-time tool: any pre-trained Gaussian scene could be stripped of never-visible Gaussians after training, which the paper only hints at as an optional export step.
If mask quality is the main quality bottleneck, then combining the background loss with mask-uncertainty weighting or interactive refinement could close most of the remaining PSNR gap.
The object-centric representation opens editing workflows like object swapping or inpainting directly in Gaussian space, which the paper does not explicitly explore.
A testable extension: apply the background loss and pruning to a newer splatting variant such as Mip-Splatting or 4D Gaussians and measure whether the same proportional savings appear.

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.

Desk Editor's Note

read the letter

Colleague, here's my read on Rogge and Stricker's Object-Centric 2DGS. The occlusion-aware pruning is the real contribution: tracking which Gaussians are actually blended, rather than just in the frustum, removes ~10% of Gaussians for free, with no measurable quality change on DTU or Mip-NeRF360. That's a clean, general idea and they validate it in both 2DGS and 3DGS. The background loss is simpler — mask the photometric loss and add an alpha penalty — but it works, cutting model size by more than 95% and training time by 70% on the large scenes. The DTU chamfer distance results are credible and show only a minor quality drop.

The soft spot is the Mip-NeRF360 evaluation. Masked PSNR and SSIM are computed with the same SAM 2 masks that produce the background loss. If those masks systematically exclude object pixels, the metric never looks at those pixels, so both the baseline and the masked model are judged only on mask-included regions. That can inflate the masked model's apparent quality and hide failures exactly where the method is most vulnerable. The paper acknowledges mask sensitivity and shows a failure case in Appendix C, but the quantitative claim still rests on a non-independent metric. The right fix is an evaluation on a manually verified object region or at least a sensitivity analysis with mask perturbations. This doesn't sink the paper, but it means the 'retaining competitive quality' line is weaker than the DTU numbers suggest.

Also, no code, no error bars, and the SSIM masking is nonstandard (they acknowledge it biases the score slightly). Those are minor but worth noting.

Overall, the paper is a solid practical contribution. The pruning trick alone is worth citing. The paper deserves a serious referee; the authors should be asked to address the evaluation circularity and release code.

Recommendation: send to peer review with a request for a revised evaluation protocol.

Referee Report

3 major / 5 minor

Summary. The paper proposes two extensions to 2D Gaussian Splatting (2DGS): (1) a background loss that uses segmentation masks to force Gaussian opacity to zero outside the target object, and (2) an occlusion-aware pruning strategy that removes Gaussians which never contribute to any rendered pixel. The method is evaluated on DTU with ground-truth chamfer distance and on Mip-NeRF360 with masked PSNR/SSIM. The authors report that pruning alone reduces Gaussian counts by roughly 10% with no measurable quality change, and that the full method reduces model size by about 96% and training time by about 71% on Mip-NeRF360, while the masked PSNR drops from 28.31 to 25.69 dB. The paper also includes ablations, a 3DGS demonstration of the pruning strategy, and a qualitative discussion of mask failure cases.

Significance. The occlusion-aware pruning strategy is simple, practical, and appears to provide a 'free' model-size reduction on standard multi-view benchmarks, with support from the DTU ground-truth geometry experiments. The background loss formulation is clean and, on DTU, the object-centric reconstruction achieves chamfer distances close to the 2DGS baseline while reducing training time and Gaussian count substantially. If the pruning result is robust, it is a generally applicable improvement for Gaussian-based methods, and the 3DGS validation in Appendix B supports that transferability. However, the headline 'competitive quality' claim on Mip-NeRF360 is weakened because the evaluation masks are the same masks used in the training objective, so systematic mask errors are not independently assessed; this is a load-bearing issue for the abstract's central claim, though it is addressable with additional evaluation.

major comments (3)

[Section 5.3.2, Table 4] The masked PSNR and SSIM on Mip-NeRF360 are computed with the same SAM 2 masks M_i that define the background loss L_b (Eq. 1) and the masked photometric loss L_M_c (Eq. 2). If M_i systematically excludes object pixels, those pixels are discarded from the metric by the division by sum(M) in Eq. 5, so the proposed method receives no penalty for missing them, and the full-scene baseline receives no credit for reconstructing them. The paper's own Appendix C documents a consistent mask failure on the Bicycle scene, and Table 4 aggregates this scene into the reported mean of 25.69 dB. To support the abstract's 'competitive quality' claim, the authors should either evaluate with independently verified masks (e.g., manually corrected or a second segmentation method) or report per-scene results together with mask error statistics (inclusion and exclusion rates). Without this, the masked metric may partially reward modeling mask artifacts rather than the object.
[Section 5.3.2, Table 4] The reported full-method quality drop on Mip-NeRF360 is 2.6 dB in masked PSNR (28.31 to 25.69), which the text describes as 'a small decrease in quality.' This is not self-evidently small in the context of novel-view synthesis benchmarks, and the magnitude of the drop is itself entangled with the mask-defined metric. The abstract's phrasing 'retaining competitive quality' should be tempered, or the authors should provide an additional quality measure that is less dependent on the training masks, such as a perceptual metric on the masked region, a geometry metric from the extracted meshes, or a manual evaluation of a subset of views.
[Section 4.3] The pruning strategy assumes that any Gaussian that is never visible in any training view is unnecessary. This is plausible for dense multi-view captures like DTU and Mip-NeRF360, but the paper states the claim generally ('without compromising quality' in the abstract and Section 4.3). For sparser view configurations, a Gaussian that is occluded in all training views might still be visible from a novel viewpoint outside the training set, and the background loss does not protect such Gaussians because they receive no gradients. The authors should add an explicit caveat about this assumption and, ideally, demonstrate on a reduced-view subset of DTU that the pruning does not degrade novel-view rendering outside the training poses.

minor comments (5)

[Abstract] The phrase 'up to 96% smaller and up to 71% faster' is ambiguous because Table 4 reports means over the Mip-NeRF360 scenes, not per-scene maxima; please clarify whether the abstract refers to average or maximum reductions.
[Section 5.3.2] The masked SSIM procedure applies the mask to the ground truth and rendered image before computing SSIM and then averages over valid pixels; this allows invalid pixels to enter the local windows. The paper acknowledges this, but a masked SSIM implementation that excludes invalid pixels from the window statistics would be more rigorous.
[Figure 8 and Section 5.4.3] The ablation of the background-loss coefficient λ is shown only qualitatively; a quantitative curve of reconstruction quality versus λ on a representative scene would make the choice of γ = 0.5 more convincing.
[Tables 4 and 5] The number of occluded Gaussians removed for 2DGS* (328,088 for outdoor scenes in Table 5) is larger than the total Gaussian reduction between 2DGS* and Ours* in Table 4 (177,879); the paper should explain that adaptive densification can add new Gaussians after pruning, as this discrepancy is otherwise confusing.
[Section 5.1] The sentence 'Although the inputs to the photometric loss are masked as described in Section 4.2.' is grammatically incomplete and should be revised.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is evaluated against external DTU ground truth and symmetric masked metrics, and no load-bearing claim reduces to a fitted input or self-citation.

full rationale

The paper's central claims—smaller models, faster training, and competitive quality—are not obtained by re-labeling fitted inputs as predictions. The background loss (Eq. 1) and masked photometric loss (Eq. 2) use segmentation masks M_i as inputs, and the Mip-NeRF360 quality numbers (Table 4) are masked PSNR/SSIM computed with the same masks (Eqs. 4–5). This raises an evaluation-validity concern: if the SAM 2 masks systematically exclude object pixels, those pixels are neither supervised during training nor counted in the metric, which could inflate the "competitive quality" claim. However, this is not circularity in the sense defined here: no parameter is fitted to the reported metric, the 2DGS baseline is evaluated under the identical masking, and the paper's central efficiency claims do not depend on those PSNR values. The DTU mesh evaluations use externally scanned ground-truth point clouds and chamfer distance, which are independent of the masks. The pruning claim is supported by explicit visibility tracking, and its "without compromising quality" property follows directly from removing only Gaussians that are never used in alpha blending; this is a definitional exactness property rather than a circular prediction. All citations (2DGS, 3DGS, SAM 2, NeUS, etc.) are external prior work; no load-bearing argument reduces to a self-citation or imported uniqueness theorem. The method's assumptions about mask quality are explicitly acknowledged in Section 5.5 and Appendix C, and those limitations concern robustness, not circularity. Overall score 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the accuracy of input masks and on the training-view coverage assumption for pruning. The method introduces no new physical entities. The only hand-chosen quantities are the background loss weight, pruning interval, and mesh voxel size.

free parameters (3)

lambda_b (gamma) = 0.5
Weight of the background loss Lb in Eq. 3. Set to 0.5 for all experiments; the paper shows a qualitative ablation (Figure 8) with small and large lambda but does not systematically tune it per scene.
Pruning interval = 100 iterations (DTU), 600 iterations (Mip-NeRF360)
Interval at which occluded Gaussian pruning runs (Section 5.1). Chosen by dataset scale, not swept.
Mesh extraction voxel size = 0.004
Voxel size for unbounded TSDF mesh extraction on Mip-NeRF360 (Section 5.1). Chosen for qualitative comparison, not tuned.

assumptions (5)

standard math The differentiable Gaussian rasterizer and its gradients (from 3DGS/2DGS) correctly model the scene's photometry and geometry.
The method builds directly on the 2DGS renderer (Section 3.2); correctness of the whole pipeline is assumed from the base works.
domain assumption Input segmentation masks Mi accurately delineate the target object across all views.
The background loss penalizes opacity outside Mi and the photometric loss is masked by Mi. The paper's own limitation section (5.5) states systematic mask errors yield bad reconstructions; Figure 5 and Appendix C show failures.
domain assumption Gaussians that are never used in alpha blending for any training view are unnecessary for novel-view rendering of the object.
The pruning strategy removes these Gaussians during training. This assumes the training views span the object's appearance; a Gaussian visible only from a held-out view would be wrongly pruned. The paper suspends densification in the second half to avoid removing temporarily occluded Gaussians (Section 5.3.2).
domain assumption Masking the photometric loss by the object mask and adding an opacity penalty on background pixels suffices to learn the object's appearance and geometry.
This is the core design choice of Section 4.2; the ablation shows it works but also that it can degrade geometry (Figure 8) when lambda is large or masks are wrong.
domain assumption The 2DGS surface-alignment regularization (depth distortion and normal consistency) remains valid on object-only Gaussian sets.
The two 2DGS regularization terms are kept unmasked in Eq. 3; the paper asserts they do not affect opacity, but their interaction with the new background loss is not analyzed in depth.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Object-Centric 2D Gaussian Splatting: Background Removal and Occlusion-Aware Pruning for Compact Object Models." pith.science (2026). https://pith.science/paper/XHLSOBF6

@misc{pith2026250108174,
  author       = {Pith},
  title        = {Pith review of: Object-Centric 2D Gaussian Splatting: Background Removal and Occlusion-Aware Pruning for Compact Object Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XHLSOBF6}},
  note         = {Machine review of arXiv:2501.08174}
}

read the original abstract

Current Gaussian Splatting approaches are effective for reconstructing entire scenes but lack the option to target specific objects, making them computationally expensive and unsuitable for object-specific applications. We propose a novel approach that leverages object masks to enable targeted reconstruction, resulting in object-centric models. Additionally, we introduce an occlusion-aware pruning strategy to minimize the number of Gaussians without compromising quality. Our method reconstructs compact object models, yielding object-centric Gaussian and mesh representations that are up to 96% smaller and up to 71% faster to train compared to the baseline while retaining competitive quality. These representations are immediately usable for downstream applications such as appearance editing and physics simulation without additional processing.

Figures

Figures reproduced from arXiv: 2501.08174 by the authors.

**Figure 1.** Our method, optimizes 2D Gaussians to accurately model specific object surfaces. They can be rendered directly or [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Overview of our method adapted from the 3DGS paper (Kerbl et al., 2023). Changes to the original pipeline are [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Mask from SAM 2 overlaid on the input image. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (6 more)

**Figure 5.** Figure 5: Example of erroneous mask overlaid on the input [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 4.** Figure 4: Qualitative comparison of rendered meshes on [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 7.** Figure 7: Visualization of occluded Gaussians for 2DGS [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: Ablation studies for background removal. Shown [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: Appearance editing on the Mip-NeRF360 dataset [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

**Figure 11.** Figure 11: Example failure case for our method with con [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

TRAN-D: 2D Gaussian Splatting-based Sparse-view Transparent Object Depth Reconstruction via Physics Simulation for Scene Update
cs.RO 2025-07 conditional novelty 6.0 of 10

TRAN-D reconstructs transparent-object depth from sparse views via segmentation-conditioned 2D Gaussian Splatting with an object-aware loss, and updates scenes after object removal using one image and physics simulation.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages · cited by 1 Pith paper

[1]

Martin-Brualla, R., and Srinivasan, P. P. (2021). Mip- nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV

work page 2021
[2]

T., Mildenhall, B., Verbin, D., Srinivasan, P

Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. (2022). Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR

work page 2022
[3]

Yao, Y . (2023). Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv:2311.16043. Gu´edon, A. and Lepetit, V . (2024). Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR

arXiv 2023
[4]

P., Mildenhall, B., Barron, J

Hedman, P., Srinivasan, P. P., Mildenhall, B., Barron, J. T., and Debevec, P. (2021). Baking neural radiance fields for real-time view synthesis. ICCV

work page 2021
[5]

Huang, B., Yu, Z., Chen, A., Geiger, A., and Gao, S. (2024). 2d gaussian splatting for geometrically accurate radi- ance fields. In SIGGRAPH 2024 Conference Papers . Association for Computing Machinery

work page 2024
[6]

Jensen, R., Dahl, A., V ogiatzis, G., Tola, E., and Aanæs, H. (2014). Large scale multi-view stereopsis evalua- tion. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406–413. IEEE

work page 2014
[7]

Kerbl, B., Kopanas, G., Leimk ¨uhler, T., and Drettakis, G. (2023). 3d gaussian splatting for real-time radi- ance field rendering. ACM Transactions on Graphics, 42(4)

work page 2023
[8]

Ramamoorthi, R., and Ng, R. (2020). Nerf: Repre- senting scenes as neural radiance fields for view syn- thesis. In ECCV

work page 2020

Show all 16 references

[9]

V ., Carion, N., Wu, C.-Y ., Girshick, R., Doll´ar, P., and Feichtenhofer, C

Mintun, E., Pan, J., Alwala, K. V ., Carion, N., Wu, C.-Y ., Girshick, R., Doll´ar, P., and Feichtenhofer, C. (2024). Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714

2024 arXiv
[10]

T., and Hedman, P

Mildenhall, B., Geiger, A., Barron, J. T., and Hedman, P. (2023). Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. SIG- GRAPH. Sch¨onberger, J. L. and Frahm, J.-M. (2016). Structure- from-motion revisited. In Conference on Computer Visio...

2023
[11]

Wang, W. (2021). Neus: Learning neural implicit sur- faces by volume rendering for multi-view reconstruc- tion. NeurIPS

2021
[12]

Jiang, C. (2023). Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198

2023 arXiv
[13]

Shen, W., and Tian, Q. (2024). Gaussianobject: High- quality 3d object reconstruction from four views with gaussian splatting. ACM Transactions on Graphics , 43(6)

2024
[14]

Yariv, L., Gu, J., Kasten, Y ., and Lipman, Y . (2021). V ol- ume rendering of neural implicit surfaces. In Thirty- Fifth Conference on Neural Information Processing Systems

2021
[15]

P., Szeliski, R., Barron, J

Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P. P., Szeliski, R., Barron, J. T., and Mildenhall, B. (2023). Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv

2023
[16]

Zhou, Q.-Y ., Park, J., and Koltun, V . (2018). Open3D: A modern library for 3D data processing. arXiv:1801.09847. APPENDIX A DOWNSTREAM APPLICA TIONS Our method produces an isolated representation of a target object from the scene. Whether using Gaus- sians or mesh, the repre...

2018 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Martin-Brualla, R., and Srinivasan, P. P. (2021). Mip- nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV

work page 2021

[2] [2]

T., Mildenhall, B., Verbin, D., Srinivasan, P

Barron, J. T., Mildenhall, B., Verbin, D., Srinivasan, P. P., and Hedman, P. (2022). Mip-nerf 360: Unbounded anti-aliased neural radiance fields. CVPR

work page 2022

[3] [3]

Yao, Y . (2023). Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv:2311.16043. Gu´edon, A. and Lepetit, V . (2024). Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR

arXiv 2023

[4] [4]

P., Mildenhall, B., Barron, J

Hedman, P., Srinivasan, P. P., Mildenhall, B., Barron, J. T., and Debevec, P. (2021). Baking neural radiance fields for real-time view synthesis. ICCV

work page 2021

[5] [5]

Huang, B., Yu, Z., Chen, A., Geiger, A., and Gao, S. (2024). 2d gaussian splatting for geometrically accurate radi- ance fields. In SIGGRAPH 2024 Conference Papers . Association for Computing Machinery

work page 2024

[6] [6]

Jensen, R., Dahl, A., V ogiatzis, G., Tola, E., and Aanæs, H. (2014). Large scale multi-view stereopsis evalua- tion. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406–413. IEEE

work page 2014

[7] [7]

Kerbl, B., Kopanas, G., Leimk ¨uhler, T., and Drettakis, G. (2023). 3d gaussian splatting for real-time radi- ance field rendering. ACM Transactions on Graphics, 42(4)

work page 2023

[8] [8]

Ramamoorthi, R., and Ng, R. (2020). Nerf: Repre- senting scenes as neural radiance fields for view syn- thesis. In ECCV

work page 2020