Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

CF3: Compact and Fast 3D Feature Fields

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A semantic 3D feature field can be compressed to 5% of prior Gaussians without losing query accuracy.

desk verdict A pragmatic top-down compression pipeline that delivers a 20-100x reduction in Gaussian footprint for 3D feature fields with competitive mIoU; deserves refereeing, but needs error bars and a better-validated variance cutoff. read the letter →

arxiv 2508.05254 v3 pith:UYYGECLB submitted 2025-08-07 cs.CV cs.AI

Hyunjoon Lee , Joonkyu Min , Jaesik Park This is my paper
classification cs.CVcs.AI
keywords 3DGaussianSplattingfeaturefieldcompressionopen-vocabularysegmentationliftingper-Gaussianautoencoderadaptivesparsificationsemanticfieldsreal-timerendering
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 a semantic 3D feature field does not need to be optimized jointly with color. Starting from an already-trained 3D Gaussian Splatting scene, CF3 fuses multi-view 2D foundation-model features into 3D with an alpha-weighted average, trains a per-Gaussian autoencoder on those lifted features to compress them to three numbers, and then prunes and merges redundant Gaussians. The claimed outcome is a feature field matching Feature-3DGS within about one or two mIoU points while using roughly 5% of its Gaussians (47k–55k versus 636k–1423k) and storing 121×–245× less (3.6–4.2MB versus 464M–1032MB). A sympathetic reader would care because it points toward real-time open-vocabulary querying on large or outdoor scenes without a separate feature representation.

What carries the argument

Two mechanisms carry the argument. The first is the weighted-sum feature-lifting identity of Eq. (4), which converts the feature-embedding problem into a closed-form $\alpha$-weighted average over the pixels each Gaussian affects; it simultaneously produces the variance estimate of Eq. (5) used to filter unreliable Gaussians. The second is adaptive sparsification: global-contribution pruning (Eq. 13) plus merging of neighbors that pass a Mahalanobis-distance threshold ($d_M < \chi^2_\beta$) and a cosine-similarity threshold, with Gaussian attributes recombined by moment-matching formulas (Eqs. 15–18). The 5-layer per-Gaussian autoencoder (encoding $[128,64,32,16,3]$) is what makes feature rende

What would settle it

Use a pretrained 3DGS scene and inject 1,000 extra semi-transparent floaters in empty space before running CF3; if downstream open-vocabulary mIoU drops sharply while rendered RGB stays nearly unchanged, the lifting stage is not robust to the pretrained geometry on which it depends. A second check: on a synthetic mirror or glass scene, compare CF3's segmentation of the reflective surface with Feature-3DGS; alpha-weighted averaging over views should blur that surface's semantics if the linear-lifting assumption is wrong.

Watch

Extended reading notes

Core claim

The central claim is that most Gaussians produced for photorealistic color rendering are redundant for semantic feature representation, and that a top-down construction can exploit that redundancy without a joint optimization loop. CF3 assigns each Gaussian a view-consistent reference feature by weighted fusion (Eq. 4): $f_i \approx \frac{\sum_{m,p} w_{i,m,p}F_{m,p}}{\sum_{m,p} w_{i,m,p}}$, where $w_{i,m,p}$ is the $\alpha$-blending weight of Gaussian $i$ at pixel $p$ in image $m$. This closed-form lifting replaces the usual per-pixel optimization for features, and it lets the paper train a per-Gaussian autoencoder directly on the 3D feature distribution rather than on 2D feature maps. The auto

Load-bearing premise

The load-bearing premise is that a Gaussian's correct feature is the alpha-weighted average of the 2D features of the pixels it influences; when the pretrained 3DGS has wrong geometry or semi-transparent floaters, this average mixes unrelated content and the autoencoder and sparsification inherit that noise.

Editorial extensions

If this is right

  • Feature fields can be built from an existing color 3DGS without re-optimizing color, so semantic embedding becomes a post-processing stage rather than part of scene reconstruction.
  • Because the compressed feature lives in the RGB channels, any 3DGS renderer that can output color can output features, making real-time open-vocabulary queries available on the same hardware.
  • Storage for semantic fields drops to single-digit megabytes (2.6–4.2MB, or 1.5–2.5MB with vector quantization), which makes large-scale scenes such as KITTI-360 tractable (6.2MB/95k Gaussians versus 3810MB/1734k).
  • Training a per-Gaussian autoencoder on lifted 3D features rather than on 2D maps reduces the train/inference distribution gap, which the paper credits for improved segmentation on low-resolution MaskCLIP features (46.9 mIoU versus 35.9 for Feature-3DGS).
  • The pruning and merging criteria are attribute-based rather than task-specific, so the same sparsification can be applied to other per-Gaussian fields.

Reading between the lines

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

  • If the color-optimized 3DGS over-populates texture-rich regions while semantics are locally uniform, the redundancy CF3 exploits is structural rather than accidental; equivalent compression should be achievable by other post-hoc top-down pipelines, not only this one.
  • The alpha-weighted average in Eq. (4) is linear, so any view-dependent semantic effect (glossy reflections, transparent objects, thin structures) will be blended away or corrupted; a testable extension is to evaluate per-category mIoU on reflective versus diffuse objects.
  • The top-0.01% variance filter only removes extreme outliers; if misreconstructed floaters are systematic, a stronger geometry-consistency prior during lifting or sparsification may be needed.
  • Because the final field stores only 3 numbers per Gaussian, it could be combined with existing attribute-compression schemes as the paper demonstrates with vector quantization; the same extra-storage argument likely extends to full color-plus-feature joint compression.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes CF3, a top-down pipeline for building compact 3D feature fields from pre-trained 3D Gaussians. In a first stage, multi-view 2D features are lifted into a per-Gaussian reference feature via a weighted sum (Eq. 4), outliers are filtered by a variance criterion, and a per-Gaussian autoencoder compresses the lifted features into a 3-dimensional latent. In a second stage, adaptive sparsification prunes and merges redundant Gaussians while optimizing Gaussian attributes against rendered reference features and depth. Experiments on Replica, LERF, 3D-OVS, and KITTI-360 compare CF3 with Feature-3DGS and LangSplat, reporting substantially fewer Gaussians (about 5--7% of Feature-3DGS), much lower storage, higher FPS, and mIoU within a few points of the baselines.

Significance. If the reported numbers hold, the paper makes a useful practical contribution: it demonstrates that color-optimized Gaussian splats are highly redundant for semantic feature fields, and that a lift-then-compress-then-sparsify pipeline can match optimization-based feature embedding at a fraction of the storage and a large speedup. The pipeline is simple, the ablations support the contribution of pruning and merging, and the use of a 3-dimensional latent directly with the 3DGS rasterizer is elegant. The main limitations are that the lifting step's noise tolerance is not quantified, one baseline's collapse with 3-dimensional features is left unexplained, and an apparently important regularization term is mentioned but not specified.

major comments (4)
  1. [§4.1, Eq. (4)] The feature-lifting approximation is load-bearing but its validity is not tested. Eq. (4) sets a Gaussian's feature to the alpha-weighted average of the pixels it influences, which is only exact for view-constant features and clean geometry. With floaters, semi-transparent boundaries, or inaccurate pretrained 3DGS geometry, a single splat can average semantically distinct features. The variance filter in §4.1 removes only the top 0.01% by variance; with ~600k Gaussians this is about 60 splats, far too few to protect against systematic geometry errors. Moreover, Table 3 row 4 (pruning+merging without variance filtering) achieves nearly the same mIoU, so the filter is not the main safeguard. The paper should include a sensitivity study: vary the variance-filter quantile, perturb the input 3DGS (e.g., fewer training iterations or random Gaussian removal), and compare lifted features against
  2. [§5.2, Tables 1–2] The paper reports that Feature-3DGS with 3-dimensional features collapses (mIoU 21.3 on Replica, 4.3 on LERF), while CF3 uses a 3-dimensional latent and achieves 70.8 and 52.4, respectively. This striking difference is never explained. Since the main novelty includes the per-Gaussian autoencoder, the reader needs to know whether the improvement comes from lifting before compression, from training the autoencoder on the lifted feature distribution, or from the decoder being trained per-Gaussian. A diagnostic would be valuable: for example, compare CF3's latent distribution with the 3-dimensional embedding learned by Feature-3DGS, or ablate the autoencoder by using a linear projection instead of the MLP. As written, the comparison raises a question about the baseline rather than establishing the mechanism of the proposed method.
  3. [§5.2, first paragraph] In the Replica/LSeg experiment, the text says 'we also incorporate the raw feature map as regularization.' This is the only mention of this term; no equation, weight, or feature source is given. If this regularization is used in the results of Table 1, then the loss in Eq. (10) is incomplete and the method is not reproducible from the paper. If it is not used, the sentence should be removed or clarified. This sentence is especially concerning because it appears exactly in the experiment that supports the main compactness claim.
  4. [Tables 1–6] All evaluation numbers appear to be single-run results with no standard deviation or repeated-seed reporting. The headline claim is 'competitive' mIoU, and in several cases the gap to Feature-3DGS is only 1--3 points (e.g., Table 1: 70.8 vs 73.4; Table 2: 52.4 vs 53.8). Without error bars, it is impossible to judge whether the degradation is statistically significant. At minimum, please report mean and std over at least three runs for the main tables, or provide per-scene results so the variance can be assessed.
minor comments (5)
  1. [§4.1, Eq. (5)] The notation 'Var(f_i)' is used as a vector of per-dimension variances, and the filter applies to its norm. This should be clarified: define Var(f_i) in R^D and the norm explicitly, since the current text could be read as a scalar variance of a vector.
  2. [§4.2] The autoencoder architecture is only partially specified: the encoder is [128, 64, 32, 16, 3], but the decoder architecture, activation functions, learning rate, and training iterations are omitted. Also, the values of λ_cos and λ_struc in Eq. (6) are not given in either the main text or the supplementary material.
  3. [Table 3] The table caption contains a formatting artifact: 'First , Second , Third Metrics' appears broken. Also, the row labels with checkmarks are hard to read; consider using explicit column names such as 'VF', 'Prune', 'Merge' and a legend.
  4. [Table 1 caption] Typo: 'Relica' should be 'Replica'.
  5. [§5.2, Tables 1–2] The baselines '3DGS*' and 'LightGaussian*' are described as 'with feature lifting' but the exact lifting protocol is not clarified (same Eq. (4)? same variance filter? same pretrained 3DGS?). Please state precisely which components of CF3 are applied to these baselines, so the comparison is interpretable.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the compact feature field is an externally evaluated compression result, not a prediction that reduces to its inputs by construction.

full rationale

The paper's central claim is a compression result: it builds a compact 3D feature field by lifting 2D features into a pretrained 3DGS via Eq. (4), training a per-Gaussian autoencoder on the lifted features (Sec. 4.2), and then pruning/merging Gaussians to reproduce the lifted feature field (Sec. 4.3). The supervision for the sparsification stage, F_ref in Eq. (11), is the lifted feature field itself, but that is the representation being compressed, not an external quantity being predicted; using the target as training signal in an autoencoder/compression pipeline is not circular. Downstream claims (mIoU, accuracy, localization) are evaluated on standard benchmarks (Replica, LERF, 3D-OVS, KITTI-360) with external text queries and ground-truth masks, so no reported metric is forced by the paper's own fitted values. Eq. (4) is explicitly acknowledged as an approximate weighted-sum lifting following prior training-free aggregation methods ([7,30]), not presented as a first-principles derivation. The only self-citation involving an author of this paper ([41] in related work) is not load-bearing for any central claim. Concerns about feature-lifting noise, the fixed 0.01% variance-filter threshold, or the absence of sensitivity analysis are robustness/correctness risks, not circularity under the definitions used here. Score 1 reflects only the minor non-load-bearing self-citation; no circular step was identified.

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

The central claim rests on the pre-trained 3DGS geometry, the validity of linear feature lifting, the sufficiency of a 3-dimensional latent bottleneck, and a set of hand-chosen pruning and merging thresholds. No new entities are introduced.

free parameters (5)
  • Autoencoder loss weights λcos, λstruc = not reported
    Balance MSE, cosine, and structure preserving losses in Eq. (6); no values or tuning procedure given.
  • Latent dimension = 3
    Chosen in Sec. 4.2 to reuse the RGB rasterizer; may limit feature discriminability.
  • Variance filter percentile = top 0.01%
    Sec. 4.1: heuristic threshold to remove Gaussians with high feature variance; no sensitivity analysis.
  • Sparsification thresholds = τcon=0.25, τsim=0.999, τgrad=1e-5, χ2β=2.38
    Supplement A: thresholds for pruning and merging; appear tuned for the evaluation datasets without a validation protocol.
  • Merge schedule = merge every 50 iterations, 3000 iterations total
    Supplement A: schedule for adaptive sparsification; no justification.
assumptions (5)
  • standard math Moment matching is a valid approximation for merging two Gaussians.
    Eqs. (15)-(18) from Gaussian mixture reduction [39]; approximate, not exact.
  • domain assumption Pre-trained 3DGS provides accurate geometry.
    Feature lifting in Sec. 4.1 assumes the existing 3DGS geometry is good enough to associate 2D features across views.
  • domain assumption Foundation model features are comparable across views.
    CLIP, LSeg, and SAM features are treated as a consistent semantic signal; multi-view inconsistency is only partially fixed by lifting.
  • ad hoc to paper Top 0.01% variance filter removes noisy Gaussians.
    Sec. 4.1: the percentile is arbitrary and no ablation is shown for it.
  • ad hoc to paper Small gradient implies a stable, well-represented region.
    Sec. 4.3 and Algorithm 1 use gradient threshold τgrad to decide when merging is safe; small gradient could also occur from a flat loss surface.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CF3: Compact and Fast 3D Feature Fields." pith.science (2026). https://pith.science/paper/UYYGECLB

@misc{pith2026250805254,
  author       = {Pith},
  title        = {Pith review of: CF3: Compact and Fast 3D Feature Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UYYGECLB}},
  note         = {Machine review of arXiv:2508.05254}
}
read the original abstract

3D Gaussian Splatting (3DGS) has begun incorporating rich information from 2D foundation models. However, most approaches rely on a bottom-up optimization process that treats raw 2D features as ground truth, incurring increased computational costs. We propose a top-down pipeline for constructing compact and fast 3D Gaussian feature fields, namely, CF3. We first perform a fast weighted fusion of multi-view 2D features with pre-trained Gaussians. This approach enables training a per-Gaussian autoencoder directly on the lifted features, instead of training autoencoders in the 2D domain. As a result, the autoencoder better aligns with the feature distribution. More importantly, we introduce an adaptive sparsification method that optimizes the Gaussian attributes of the feature field while pruning and merging the redundant Gaussians, constructing an efficient representation with preserved geometric details. Our approach achieves a competitive 3D feature field using as little as 5% of the Gaussians compared to Feature-3DGS.

Figures

Figures reproduced from arXiv: 2508.05254 by the authors.

Figure 1
Figure 1. We propose CF3 for constructing a compact and fast 3D feature field from 3D Gaussians. The previous method (Feature-3DGS) jointly optimizes features with colors, resulting in excessive Gaussians for rendering the feature field. CF3 effectively compresses and sparsifies the 3D feature field while maintaining sufficient details as shown in the rendered feature maps. Abstract 3D Gaussian Splatting (3DGS) has begun inco… view at source ↗
Figure 2
Figure 2. Overview of our CF3 pipeline. We utilize pre-trained 3D Gaussians to construct a 3D feature field. We adopt a weighted-sum strategy to lift features extracted from a visual foundation model into 3D. Subsequently, a per-Gaussian autoencoder compresses high￾dimensional features into lower-dimensional embeddings, effectively removing noisy features through a variance filtering step. Afterward, adaptive sparsification m… view at source ↗
Figure 3
Figure 3. Feature lifting. The raw features from visual foundation models are not view-consistent. Feature lifting (Sec. 4.1) alleviates this inconsistency. This idea appears in recent training-free feature aggregation methods [7, 30]. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of our adaptive sparsification pipeline. Unlike the original 3D Gaussian Splatting, which preserves fine-grained details for photorealistic rendering, our method focuses on feature field reconstruction and merges redundant Gaussians to reduce unnecessary densi…
Figure 5
Figure 5. Figure 5: Qualitative comparison. We visualize open-vocabulary semantic segmentation and localization results using CLIP [36] with SAM [20] features on the LERF [18] dataset. Our method shows precise results even for small objects in these tasks. Feature-3DGS [56] is tested with…
Figure 6
Figure 6. Figure 6: Qualitative comparison. We visualize open-vocabulary semantic segmentation results using MaskCLIP [55] features. Feature￾3DGS [56] is tested with the speed-up module (128 dim). We highlight the ground truth masks in red for the corresponding query texts. Config MaskCLI…
Figure 8
Figure 8. Figure 8: Additional Result on KITTI-360 Dataset. We visualize each Gaussian in CF3 based on its similarity to the text query and render the result. Blue indicates low similarity, while higher similarity is shown in red. Storage↓ FPS↑ #G↓ 3DGS* 3810.2M 1.8 1734k CF3 (Ours) 6.2M …
Figure 7
Figure 7. Figure 7: 3D Segmentation Results. We perform open-vocabulary 3D segmentation on the 3D-OVS dataset. The following queries are used, in order: a book of The Unbearable Lightness of Being, Coca￾Cola, a red Nintendo Switch Joy-Con controller, and Dinosaur. Storage↓ FPS↑ mIoU↑ #G↓ …

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. CoSAG: Compact Semantic Anchor Gaussians via Training-Free Rate-Distortion Coding

    cs.CV 2026-07 accept novelty 6.0 of 10

    Training-free closed-form lift plus spatially predictive entropy coding of Gaussian-to-anchor bindings yields sub-megabyte open-vocabulary 3D fields that match or beat prior accuracy.

  2. NRGS: Neural Regularization for Robust 3D Semantic Gaussian Splatting

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    A variance-aware conditional MLP operating on 3D Gaussians corrects semantic errors from multi-view inconsistent 2D features to produce more accurate and robust 3D semantic Gaussian Splatting.

  3. C3G: Learning Compact 3D Representations with 2K Gaussians

    cs.CV 2025-12 unverdicted novelty 6.0 of 10

    C3G creates compact 3D Gaussian representations with 2K points by guiding placement via learnable tokens that aggregate multi-view features through attention, yielding better efficiency and performance than dense methods.

Reference graph

Works this paper leans on

57 extracted references · 40 canonical work pages · cited by 3 Pith papers

  1. [1]

    Milena T Bagdasarian, Paul Knoll, Yi-Hsin Li, Florian Barthel, Anna Hilsmann, Peter Eisert, and Wieland Mor- genstern. 3dgs. zip: A survey on 3d gaussian splatting com- pression methods. arXiv preprint arXiv:2407.09510, 2024. 4

    arXiv 2024

  2. [2]

    M. T. Bagdasarian, P. Knoll, Y . Li, F. Barthel, A. Hils- mann, P. Eisert, and W. Morgenstern. 3DGS.zip: A Sur- vey on 3D Gaussian Splatting Compression Methods. Com- puter Graphics Forum, page e70078, 2025. https://w- m.github.io/3dgs-compression-survey/. 4

    work page 2025

  3. [3]

    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 recog- nition, pages 5470–5479, 2022. 4

    work page 2022

  4. [4]

    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, pages 422–438. Springer, 2024. 3, 4

    work page 2024

  5. [5]

    Feat2gs: Probing visual foundation models with gaussian splatting

    Yue Chen, Xingyu Chen, Anpei Chen, Gerard Pons-Moll, and Yuliang Xiu. Feat2gs: Probing visual foundation models with gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6348–6361, 2025. 2

    work page 2025

  6. [6]

    Hac++: Towards 100x compression of 3d gaussian splatting

    Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. Hac++: Towards 100x compression of 3d gaussian splatting. arXiv preprint arXiv:2501.12255, 2025. 4

    arXiv 2025

  7. [7]

    Occam’s lgs: A simple approach for lan- guage gaussian splatting

    Jiahuan Cheng, Jan-Nico Zaech, Luc Van Gool, and Danda Pani Paudel. Occam’s lgs: A simple approach for lan- guage gaussian splatting. arXiv preprint arXiv:2412.01807,

    arXiv

  8. [8]

    Gaussianpro: 3d gaussian splatting with progressive propagation

    Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty- first International Conference on Machine Learning , 2024. 3

    work page 2024

Show all 57 references
  1. [9]

    Probing the 3d awareness of visual foundation models

    Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Ab- hishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. InProceedings of the IEEE/CVF Conference on Computer Visio...

    2024

  2. [10]

    Open- set 3d scene segmentation with rendered novel views

    Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Open- set 3d scene segmentation with rendered novel views. 2023. 3

    2023

  3. [11]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.arXiv preprint arXiv:2311.17245, 2023

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps.arXiv preprint arXiv:2311.17245, 2023. 3, 4, 5, 6

    2023 arXiv

  4. [12]

    Large spatial model: End-to-end un- posed images to semantic 3d.Advances in Neural Information Processing Systems, 37:40212–40229, 2025

    Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end un- posed images to semantic 3d.Advances in Neural Information Processing Systems, 37:40212–40229, 2025. 3

    2025

  5. [13]

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

    Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. In European Conference on Computer Vision, pages 165–181. Springer,

  6. [14]

    Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation

    Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1–11. IEEE, 2022. 3

    2022

  7. [15]

    Scaling open-vocabulary image segmentation with image-level labels

    Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540–557. Springer, 2022. 3

    2022

  8. [16]

    Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings

    Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Ea- gles: Efficient accelerated 3d gaussians with lightweight en- codings. In European Conference on Computer Vision, pages 54–71. Springer, 2024. 3

    2024

  9. [17]

    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 Trans. Graph., 42(4):139–1, 2023. 3, 5

    2023

  10. [18]

    Lerf: Language embedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 19729–19739,

  11. [19]

    3d gaussian splatting as markov chain monte carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Jeff Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. arXiv preprint arXiv:2404.09591,

    arXiv

  12. [20]

    Segment any- thing

    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 any- thing. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 202...

    2023

  13. [21]

    Decomposing nerf for editing via feature field distillation

    Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Ad- vances in Neural Information Processing Systems, 35:23311– 23330, 2022. 3

    2022

  14. [22]

    Panoptic neural fields: A semantic object-aware neural scene representation

    Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Car- oline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on...

    2022

  15. [23]

    Compact 3d gaussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 3

    2024

  16. [24]

    Compression of 3d gaussian splatting with optimized feature planes and standard video codecs

    Soonbin Lee, Fangwen Shu, Yago Sanchez, Thomas Schierl, and Cornelius Hellge. Compression of 3d gaussian splatting with optimized feature planes and standard video codecs. arXiv preprint arXiv:2501.03399, 2025. 4

    2025 arXiv

  17. [25]

    Language-driven semantic seg- mentation

    Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Ren ´e Ranftl. Language-driven semantic seg- mentation. arXiv preprint arXiv:2201.03546, 2022. 1, 3, 5, 6

    2022 arXiv

  18. [26]

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

    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. arXiv preprint arXiv:2412.10231, 2024. 2, 3

    2024

  19. [27]

    Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d

    Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292–3310, 2022. 8

    2022

  20. [28]

    Weakly supervised 3d open- vocabulary segmentation

    Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open- vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433–53456, 2023. 3, 8

    2023

  21. [29]

    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. 3

    2024

  22. [30]

    Ludvig: Learning-free uplifting of 2d visual features to gaussian splatting scenes

    Juliette Marrie, Romain M ´en´egaux, Michael Arbel, Diane Larlus, and Julien Mairal. Ludvig: Learning-free uplifting of 2d visual features to gaussian splatting scenes. arXiv preprint arXiv:2410.14462, 2024. 3, 4

    2024 arXiv

  23. [31]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 1

    2021

  24. [32]

    Compact 3d scene representation via self- organizing gaussian grids

    Wieland Morgenstern, Florian Barthel, Anna Hilsmann, and Peter Eisert. Compact 3d scene representation via self- organizing gaussian grids. In European Conference on Com- puter Vision, pages 18–34. Springer, 2024. 3

    2024

  25. [33]

    Compact3d: Com- pressing gaussian splat radiance field models with vector quantization

    K Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compact3d: Com- pressing gaussian splat radiance field models with vector quantization. arXiv preprint arXiv:2311.18159, 2(3), 2023. 4

    2023 arXiv

  26. [34]

    Compressed 3d gaussian splatting for accelerated novel view synthesis

    Simon Niedermayr, Josef Stumpfegger, and R ¨udiger West- ermann. Compressed 3d gaussian splatting for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10349–10358, 2024. 3

    2024

  27. [35]

    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, 3, 4, 6, 8

    2024

  28. [36]

    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

  29. [37]

    Vi- sion transformers for dense prediction

    Ren´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 7

    2021

  30. [38]

    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. 3

    2024 arXiv

  31. [39]

    Gaussian mixture reduction via clustering

    Dennis Schieferdecker and Marco F Huber. Gaussian mixture reduction via clustering. In 2009 12th international confer- ence on information fusion, pages 1536–1543. IEEE, 2009. 5

    2009

  32. [40]

    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 Con- ference on Computer Vision and Pattern Recognition, pages 5333–5343, 2024. 2, 3

    2024

  33. [41]

    Locality- aware gaussian compression for fast and high-quality render- ing

    Seungjoo Shin, Jaesik Park, and Sunghyun Cho. Locality- aware gaussian compression for fast and high-quality render- ing. In Proceedings of the Int. Conf. on Learning Representa- tions (ICLR), 2025. 3

    2025

  34. [42]

    Panoptic lifting for 3d scene understanding with neural fields

    Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bul ´o, Nor- man M ¨uller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, p...

    2023

  35. [43]

    The replica dataset: A digital replica of indoor spaces

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797,

    1906 arXiv

  36. [44]

    Neural feature fusion fields: 3d distillation of self- supervised 2d image representations

    Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self- supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443–453. IEEE, 2022. 3

    2022

  37. [45]

    Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images

    Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 3

    2022 arXiv

  38. [46]

    End-to-end rate-distortion optimized 3d gaussian representation

    Henan Wang, Hanxin Zhu, Tianyu He, Runsen Feng, Jiajun Deng, Jiang Bian, and Zhibo Chen. End-to-end rate-distortion optimized 3d gaussian representation. In European Confer- ence on Computer Vision, pages 76–92. Springer, 2024. 3

    2024

  39. [47]

    Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs

    Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2412.16932, 2024. 3

    2024 arXiv

  40. [48]

    Opengaussian: Towards point-level 3d gaussian-based open vocabulary understanding

    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. arXiv preprint arXiv:2406.02058, 2024. 2, 3

    2024 arXiv

  41. [49]

    Mesongs: Post- training compression of 3d gaussians via efficient attribute transformation

    Shuzhao Xie, Weixiang Zhang, Chen Tang, Yunpeng Bai, Rongwei Lu, Shijia Ge, and Zhi Wang. Mesongs: Post- training compression of 3d gaussians via efficient attribute transformation. In European Conference on Computer Vision, pages 434–452. Springer, 2024. 3

    2024

  42. [50]

    Improving 2D Feature Representa- tions by 3D-Aware Fine-Tuning

    Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D Feature Representa- tions by 3D-Aware Fine-Tuning. In European Conference on Computer Vision (ECCV), 2024. 2, 4

    2024

  43. [51]

    Nerflets: Local radiance fields for efficient structure-aware 3d scene represen- tation from 2d supervision

    Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene represen- tation from 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2023

  44. [52]

    Condense: Consistent 2d/3d pre- training for dense and sparse features from multi-view images

    Xiaoshuai Zhang, Zhicheng Wang, Howard Zhou, Soham Ghosh, Danushen Gnanapragasam, Varun Jampani, Hao Su, and Leonidas Guibas. Condense: Consistent 2d/3d pre- training for dense and sparse features from multi-view images. In European Conference on Computer Vision, pages 19–38. ...

    2024

  45. [53]

    Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting

    Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Heng- shuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In European Conference on Computer Vision, pages 326–342. Springer, 2024. 3

    2024

  46. [54]

    In-place scene labelling and understanding with implicit scene representation

    Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and An- drew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 15838–15847, 2021. 3

    2021

  47. [55]

    Extract free dense labels from clip

    Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696–712. Springer, 2022. 5, 7

    2022

  48. [56]

    Feature 3dgs: Supercharging 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: Supercharging 3d gaussian splatting to enable distilled feature fields. In Pro- ceedings of the IEEE/CVF Conference on Comp...

    2024

  49. [57]

    Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding

    Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. Inter- national Journal of Computer Vision, pages 1–17, 2024. 2, 3 CF3: Compact and Fast 3D Feature Fields Supplemen...

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.