Pith. sign in

REVIEW 4 major objections 5 minor 53 references

This paper proves that view-dependent color in 3D Gaussian Splatting, specifically the Spherical Harmonic coefficients up to degree 2, transforms under rotation exactly like 3x3 matrices under conjugation, and builds a rigid-body-equivarian

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 23:01 UTC pith:BHWFRGYR

load-bearing objection The SH(0–2)-to-gl(3) intertwiner is real and useful; the paper oversells SE(3) and the action-conditioned equivariance claim doesn't survive contact with the action representation. the 4 major comments →

arxiv 2607.15536 v1 pith:BHWFRGYR submitted 2026-07-17 cs.CV cs.LG

E3DGS: Unified Geometric-Photometric Equivariance for 3D Gaussian Splatting via Color-as-Geometry Embedding

Chankyo Kim , Maani Ghaffari This is my paper
classification cs.CV cs.LG
keywords 3D Gaussian SplattingSE(3) equivariancespherical harmonicsmatrix conjugationadjoint actiongl(3) carrierself-supervised learningrobot manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

3D Gaussian Splatting couples geometry (mean, covariance) with view-dependent appearance (Spherical Harmonic coefficients), but these rotate by incompatible rules, so most 3DGS learning pipelines discard all but the static part of color. This paper shows that for SH degrees 0-2 the Wigner-D rotation action is exactly a matrix-conjugation action on gl(3), the space of 3x3 matrices. It gives explicit intertwiners that lift SH coefficients into matrix slots, and lifts position and covariance into the same carrier, so every rotation-active Gaussian attribute obeys one law: M goes to R M R^T. The resulting architecture, E3DGS, processes Gaussians with this unified rule, and the paper reports that the model stays accurate under arbitrary test-time rotations, learns from fewer labels, and accelerates dynamics learning for manipulation, all without Clebsch-Gordan tensor products.

Core claim

The core discovery is an exact representation-theoretic identification of photometry with geometry. Theorem 1 constructs linear equivariant maps placing SH degrees 0, 1, 2 into the scalar, skew-symmetric, and symmetric-trace-free subspaces of gl(3), so that the Wigner-D action becomes conjugation: Phi(D(R)f) = R Phi(f) R^T. The same carrier is used for centered positions (hat map) and covariances (matrix logarithm), so all rotation-active Gaussian attributes transform by one law. Theorem 2 extends the principle: every SH degree is realized exactly as conjugation on a larger carrier End(V_k), with the 3x3 case being the smallest member. The paper then proves that an encoder built from Ad-equi

What carries the argument

The load-bearing object is the conjugation action of SO(3) on gl(3), the space of 3x3 matrices, whose irreducible decomposition matches the SH bands 0-2. The central identity is the intertwiner that rewrites the Wigner-D rotation of SH coefficients as matrix conjugation; position (via the hat map) and covariance (via the matrix logarithm) lift into the same carrier. Supporting machinery: the modified Killing form supplies invariant attention weights and gates, Frobenius-norm LayerNorm preserves orientation, and the ReLN (reductive Lie neuron) family of Ad-equivariant operations, including linear mixing, gated activation, attention, and layer norm, keeps the encoder rotation-equivariant. A ge

Load-bearing premise

The end-to-end equivariance guarantee depends on the exact Ad-equivariance of the matrix-processing primitives (linear mixing, activations, attention, layer norm) as defined in a separate construction; the paper does not re-derive these building blocks, and if any of them fails to be exactly conjugation-equivariant when combined with the gating and normalization used here, the unified equivariance claim collapses.

What would settle it

Take a trained E3DGS model, rotate an input scene by an arbitrary rotation R, and compare the output matrix channels with R times the unrotated output times R^T; also check that the invariant readout is unchanged. Any disagreement beyond floating-point precision would refute the claimed encoder equivariance.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • 3DGS-based recognition and self-supervised pretraining can now use full SH(0-2) appearance without breaking rotation equivariance, removing the standard compromise of discarding view-dependent color.
  • Rotation robustness at test time is achieved by construction, so models no longer need pose augmentation to survive camera-frame changes.
  • The elimination of Clebsch-Gordan tensor products reduces per-pair attention cost to O(1) with an invariant similarity, making equivariance more scalable for scenes with millions of Gaussians.
  • The general matrix-carrier theorem provides a recipe for extending the same color-as-geometry embedding to higher SH degrees, such as degree 3 in a 5x5 carrier, as 3DGS rendering gains photometric bandwidth.
  • The reported label-efficiency gains indicate that the unified carrier acts as a structural inductive bias that prevents overfitting to pose-correlated appearance when labels are scarce.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same lifting could absorb other SH-transformable features, such as language-aligned or semantic attributes on Gaussians, preserving equivariance; this extension is natural but not tested in the paper.
  • Because the general carrier theorem proves SH3 can be handled by a 5x5 matrix, benchmarks with degree-3 appearance could keep exact equivariance, but whether the empirical gains persist at that bandwidth remains unknown.
  • With O(1) per-pair attention cost, the design points toward equivariant processing of very large Gaussian scenes, while the paper only demonstrates object-scale and manipulation-scale settings.
  • A stress test of the assumed exactness of the imported matrix primitives, for example measuring how rotational error accumulates through many layers, would locate the practical limit of the end-to-end guarantee.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. E3DGS proposes a unified SO(3)-equivariant representation for 3D Gaussian Splatting primitives. The central theoretical result (Theorem 1) constructs explicit linear intertwiners Φℓ that map spherical-harmonic coefficients of degree ℓ≤2 into gl(3) so that the Wigner-D action becomes matrix conjugation M↦RMR^T; Theorem 2 generalizes this to arbitrary SH degree using carriers End(V_k). The paper then builds an encoder from ReLN-style Ad-equivariant operations and validates it in two regimes: a Gaussian masked autoencoder for object recognition and segmentation, and an action-conditioned Gaussian world model for RLBench manipulation. The paper claims SE(3) equivariance while avoiding Clebsch–Gordan tensor products.

Significance. The intertwiner construction is elegant and potentially useful: it gives a parameter-free, numerically verified (ϵ≈1e-15) way to treat view-dependent SH appearance as a geometric tensor in 3DGS learning, and the general carrier theorem is a clean theoretical contribution. The explicit basis conventions and numerical checks are strengths. However, the empirical validation is under-powered (single runs, no code), and the action-conditioned world-model section contains a formal gap in how action orientations are lifted. The central geometric idea is worth publishing if these issues are addressed.

major comments (4)
  1. [§5.3, Eq. (12)] The action-conditioned world model is claimed to be SO(3)-equivariant because the lifted action channel H_act=[bvt,bωt] transforms by Ad_R. This requires ωt↦Rωt under a scene rotation. For an absolute orientation quaternion, this is false: an orientation matrix O transforms as O↦RO under the paper's active rotation convention, and the axis-angle vector of RO is not R times the axis-angle vector of O. Example: R=Rz(π/2), O=Rx(π/2); Log(RO) is a 120° rotation about (1,1,1)/√3, while R·Log(O) is a 90° rotation about y. Hence H_act does not transform by Ad_R, and the equivariant dynamics guarantee f_eq_wm([H_state,H_act])↦Ad_R(...) does not follow. The RLBench results may still hold, but they cannot be attributed to exact equivariance as written. Please either represent the action rotation as a relative rotation (or angular velocity), for which conjugation is the correct law, or revise the e
  2. [§3.4/§4.1.3, Theorem 3] Theorem 3 asserts exact equivariance of the full encoder based on 'Ad-equivariant ReLN primitives', but the proofs for ReLN-Linear and ReLN-ReLU are not contained in this manuscript; they are imported from the authors' own preprint [21]. The exactness of the end-to-end guarantee depends on these primitives being Ad-equivariant when composed with the modified Killing-form gating and Frobenius-norm LayerNorm used here. Please include explicit definitions and proofs (or precise theorems from [21]) in an appendix so the central equivariance claim is self-contained.
  3. [Tables 1–3, 5; §F.3] All experimental results are single-run and reported without error bars or seed information. Table 3 is particularly affected: with 25 episodes per task, the 5.6-point average improvement over ManiGaussian (50.4 vs 44.8) is within typical run-to-run variance in RLBench unless multiple seeds are reported. The identical ID/ID and ID/SE(3) accuracies in Tables 1–2 are consistent with exact invariance, but the manuscript should state whether evaluation is deterministic and provide variance over seeds. No code is released, making the numerical verification and experimental tables non-reproducible. Please provide mean±std over at least 3 seeds for the main tables and release code or a detailed reproducibility description.
  4. [Title/Abstract vs. §3.1] The paper claims SE(3) equivariance in the title and abstract, but all formal statements (Theorems 1–4, Eq. (8)) concern SO(3). Translation is handled by centering for recognition (§3.1), but for the action-conditioned world model in §5.3 no translation-equivariance statement or proof is given for Eq. (12). Please state precisely the symmetry group and the corresponding action on states and actions for each application, and either prove translation equivariance for the dynamics or qualify the SE(3) claim.
minor comments (5)
  1. [Appendix A.3] 'Fix a SH basis ... whose coefficient vector is identified with the coefficients in this basis' is circular; the actual mapping is given only later in Eq. (26). Move the explicit convention before the proof.
  2. [Table 3] Formatting of the ManiGaussian row makes values hard to parse (e.g., 'ManiGaussian28.076.0 64.0...'). Please add spacing and align columns.
  3. [Figure 3] Axes are unlabeled (rotation angles in degrees; similarity values). Please add labels and a colorbar.
  4. [References] Reference [5] appears truncated. Reference [21] is a preprint; since it is load-bearing for Theorem 3, its status should be flagged and the relevant definitions summarized.
  5. [Section 2.3/Conclusion] The phrase 'first formulation that makes view-dependent appearance an equivariant geometric object' should be qualified relative to TensorNet [39], which already used the gl(3) carrier for Cartesian tensors; the novelty is the photometric identification, and this should be stated precisely.

Circularity Check

0 steps flagged

No significant circularity: the SH-to-gl(3) intertwiner is explicit and externally verifiable; the ReLN self-citation is dependency, not circularity.

full rationale

The central derivation is not circular. The SH(0–2)→gl(3) intertwiners are constructed explicitly in Appendix A with a fixed real-SH convention (Eqs. 21–28), and the equivariance identity (Eq. 2 / Eq. 31) is numerically checked to ≈10^{-15} (Eq. 29). The carrier decomposition gl(3) ≃ V0⊕V1⊕V2 is a standard Clebsch–Gordan fact supported by external textbook references [45,11], and the V3 impossibility check in Appendix C (Eq. 39) independently confirms the capacity claim. No parameter is fitted and then renamed as a prediction; the empirical robustness results follow from an architectural symmetry rather than from fitting the test distribution. The only self-citation is [21] for the ReLN-Linear/ReLN-ReLU primitives reused in the backbone (Sec. 4.1.3, Theorem 3/4). This is reuse/dependency, not circularity: the Ad-equivariance of those primitives is a separately checkable property and does not assume Theorem 1 or the SH-to-matrix identification. The paper also openly scopes itself to SH(0–2) and defers higher carriers (Conclusion), so no hidden input–output equivalence is present. Separately, Section 5.3 contains a formal gap that is not circularity: it claims “under a global rotation of the scene frame, both the translational command vt and the axis-angle vector ωt transform as vt ↦ Rvt and ωt ↦ Rωt,” but an absolute orientation’s axis-angle vector does not transform as Rω under the active convention. This affects the SE(3)-equivariant dynamics guarantee as a correctness issue, but it is not an instance of an output reducing to its input by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 6 axioms · 0 invented entities

The core mathematical claim is classical representation theory made explicit, with no new physical entities. Free parameters are limited to inherited training hyperparameters. The main axiomatic weight falls on the correctness of the ReLN primitives and on matching the 3DGS SH convention.

free parameters (1)
  • World-model loss weights λ_Geo, λ_Sem, λ_Dyna (inherited from ManiGaussian) = not reported
    Used in Eq. (57) for the RLBench world-model objective; values are taken from the baseline and not stated, so the exact training objective is not reproducible from the paper alone.
axioms (6)
  • standard math Wigner-D matrices D^ℓ(R) define an SO(3) irreducible representation on the space of degree-ℓ real spherical-harmonic coefficients.
    Invoked in Sec. 3.1 and Theorem 1; textbook result [45,11].
  • standard math Clebsch–Gordan decomposition V_k⊗V_k ≅ ⊕_{ℓ=0}^{2k}V_ℓ, and V_1⊗V_1 ≅ V_0⊕V_1⊕V_2.
    Basis of Eq. (1) and Theorem 2; used to argue that gl(3) carries exactly SH(0–2) and larger carriers cover higher degrees.
  • domain assumption The real-SH basis used by 3DGS renderers is orthonormal and matches the ordering/normalization conventions in Appendix A.
    The intertwiners in Appendix A are tied to a specific real-SH convention; the paper states that another convention changes the constants by a fixed orthogonal change of basis but does not test multiple conventions.
  • domain assumption ReLN-Linear and ReLN-ReLU from Kim et al. [21] are exactly Ad-equivariant.
    The E3DGS backbone reuses these operations (Sec. 4.1); they are not re-proven here, so the end-to-end equivariance guarantee inherits its correctness from [21].
  • standard math Matrix logarithm commutes with orthogonal congruence, and the hat map satisfies [Rv]× = R[v]×Rᵀ.
    Used for the covariance and position lifts in Sec. 3.3 and Appendix D.2.
  • domain assumption Centering or relative coordinates fully handle translations, so the construction is SE(3) rather than merely SO(3).
    Asserted in Sec. 3.1 and D.5 for recognition and dynamics; exact translation behavior for the dynamics head is not formally proven.

pith-pipeline@v1.3.0-alltime-deepseek · 20897 in / 15261 out tokens · 160948 ms · 2026-08-01T23:01:05.244104+00:00 · methodology

0 comments
read the original abstract

3D Gaussian Splatting (3DGS) captures scenes by coupling explicit geometry (position, covariance) with view-dependent photometry (Spherical Harmonics). However, building $\mathrm{SE}(3)$-equivariant architectures on these primitives presents a fundamental representation bottleneck. Color has been treated as a signal rather than a geometric entity, making it nontrivial to unify symmetry across geometry and appearance as the camera frame changes. While translations are handled by relative coordinates, rotations act heterogeneously across attributes: $\mu\mapsto R\mu$, $\Sigma\mapsto R\Sigma R^\top$, and $f_\ell\mapsto D^\ell(R)f_\ell$. This mismatch complicates strict equivariance, leading existing methods to either discard or flatten SH coefficients, thereby breaking symmetry. We propose a unified solution rooted in representation theory: for SH degrees $\ell\le2$, photometry is algebraically isomorphic to a rank-2 geometric tensor. We prove that the Wigner-$D$ action on these SH coefficients can be exactly reformulated as the conjugation action on $3\times3$ matrices. Leveraging this, we introduce the Unified Matrix Embedding, a lifting that maps all Gaussian attributes into a unified carrier space, $\mathfrak{gl}(3)$. Building on the "Color-as-Geometry" formulation, we present E3DGS, a rigid-body ($\mathrm{SE}(3)$) equivariant architecture that processes 3D Gaussians without Clebsch-Gordan tensor products. Evaluations on object vision and action-conditioned Gaussian world modeling demonstrate that our unified approach yields strong robustness under camera-frame changes and improved data efficiency.

Figures

Figures reproduced from arXiv: 2607.15536 by Chankyo Kim, Maani Ghaffari.

Figure 1
Figure 1. Figure 1: Overview of E3DGS. 3D Gaussian primitives couple geometry (mean/covariance) and view-dependent photometry (SH), which transform under camera-frame changes by heterogeneous SO(3) rules. E3DGS lifts both into a unified matrix carrier and processes them with a rigid-body equivariant backbone, enabling robust object recognition and action-conditioned Gaussian world modeling. rotation poses a severe challenge. … view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the Base E3DGS Gaussian Masked Autoencoder. We instantiate our framework as a Gaussian MAE for object vision tasks (Section 5.1). Input 3D Gaussians undergo Unified Matrix Lifting to embed geometry and photometry into a single carrier. Following Gaussian Feature Grouping, scalar invariants modulate the equivariant features via invariant gating (⊗). The tokens are partially masked, processed… view at source ↗
Figure 3
Figure 3. Figure 3: Pooled invariant descriptor similarity across input rotations. Pairwise cosine similarities of pooled invariant descriptors obtained after B˜-based contractions/readout for a test instance rotated 0◦–360◦ . E3DGS (left) remains stable after invariant readout, whereas G-MAE (right) exhibits severe representational drift. is quantitatively supported by consistently lower Chamfer Distance (CD) scores (inset i… view at source ↗
Figure 4
Figure 4. Figure 4: Data-efficiency. E3DGS￾MAE outperforms the baseline across restricted training subsets. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison under out-of-distribution transformations. Under test￾time rotations (ID/SE(3)), the non-equivariant baseline (G-MAE) suffers severe representation collapse, exhibiting color corruption, missing topology, and geometric distortion (orange dashed boxes). Conversely, our equivariant E3DGS-MAE preserves structural fidelity and view-dependent appearance regardless of input orientation. In… view at source ↗
Figure 6
Figure 6. Figure 6: Dynamic Loss. E3DGS ac￾celerates the learning of rigid-body state transitions, converging faster to a lower steady-state loss than the baseline. 6 Conclusion We presented E3DGS, a framework introducing rigid-body (SE(3)) equivariance to 3D Gaussian Splatting. At its core is the Unified Matrix Lifting, which jointly embeds spatial geometry and view-dependent photometry into a single gl(3) carrier. By identi… view at source ↗
Figure 7
Figure 7. Figure 7: Masked pretraining dynamics on 3D Gaussian splats. Comparison of convergence between the baseline Gaussian-MAE [28] and E3DGS over 300 pretraining epochs on ShapeSplat, using Gaussian inputs with SH coefficients up to degree 2. E3DGS shows faster early convergence and consistently lower loss across all tracked metrics, including rotation, scale, density, Chamfer distance, spherical harmonics, and total tra… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    arXiv preprint arXiv:2206.04176 (2022)

    Assaad, S., Downey, C., Al-Rfou, R., Nayakanti, N., Sapp, B.: VN-Transformer: Rotation- Equivariant Attention for Vector Neurons. arXiv preprint arXiv:2206.04176 (2022)

    Pith/arXiv arXiv 2022

  2. [2]

    NeurIPS36, 55260–55284 (2023) 15

    Batatia, I., Geiger, M., Munoz, J., Smidt, T., Silberman, L., Ortner, C.: A General Framework for Equivariant Neural Networks on Reductive Lie Groups. NeurIPS36, 55260–55284 (2023) 15

    2023

  3. [3]

    NeurIPS35, 11423–11436 (2022)

    Batatia, I., Kovacs, D.P., Simm, G., Ortner, C., Cs´ anyi, G.: MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. NeurIPS35, 11423–11436 (2022)

    2022

  4. [4]

    arXiv preprint arXiv:2104.13478 (2021)

    Bronstein, M.M., Bruna, J., Cohen, T., Veliˇ ckovi´ c, P.: Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges. arXiv preprint arXiv:2104.13478 (2021)

    Pith/arXiv arXiv 2021

  5. [5]

    Cen, J., Fang, J., Yang, C., Xie, L., Zhang, X., Shen, W., Tian, Q.: Segment Any 3D Gaussians 39(2), 1971–1979 (2025)

    1971

  6. [6]

    19457–19467 (2024)

    Charatan, D., Li, S.L., Tagliasacchi, A., Sitzmann, V.: pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction pp. 19457–19467 (2024)

    2024

  7. [7]

    289–305 (2024)

    Choi, S., Song, H., Kim, J., Kim, T., Do, H.: Click-Gaussian: Interactive Segmentation to Any 3D Gaussians pp. 289–305 (2024)

    2024

  8. [8]

    In: ICML

    Cohen, T., Welling, M.: Group Equivariant Convolutional Networks. In: ICML. pp. 2990–2999. PMLR (2016)

    2016

  9. [9]

    Cohen, T.S., Geiger, M., K¨ ohler, J., Welling, M.: Spherical CNNs (2018)

    2018

  10. [10]

    12200–12209 (2021)

    Deng, C., Litany, O., Duan, Y., Poulenard, A., Tagliasacchi, A., Guibas, L.J.: Vector Neurons: A General Framework for SO(3)-Equivariant Networks pp. 12200–12209 (2021)

    2021

  11. [11]

    Edmonds, A.R.: Angular momentum in quantum mechanics, vol. 4. Princeton university press (1996)

    1996

  12. [12]

    3165–3176 (2020)

    Finzi, M., Stanton, S., Izmailov, P., Wilson, A.G.: Generalizing Convolutional Neural Networks for Equivariance to Lie Groups on Arbitrary Continuous Data pp. 3165–3176 (2020)

    2020

  13. [13]

    In: International conference on machine learning

    Finzi, M., Welling, M., Wilson, A.G.: A Practical Method for Constructing Equivariant Multilayer Perceptrons for Arbitrary Matrix Groups. In: International conference on machine learning. pp. 3318–3328. PMLR (2021)

    2021

  14. [14]

    NeurIPS33, 1970–1981 (2020)

    Fuchs, F., Worrall, D., Fischer, V., Welling, M.: SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks. NeurIPS33, 1970–1981 (2020)

    1970

  15. [15]

    arXiv preprint arXiv:2207.09453 (2022)

    Geiger, M., Smidt, T.: e3nn: Euclidean Neural Networks. arXiv preprint arXiv:2207.09453 (2022)

    Pith/arXiv arXiv 2022

  16. [16]

    arXiv preprint arXiv:2403.15624 (2024)

    Guo, J., Ma, X., Fan, Y., Liu, H., Li, Q.: Semantic Gaussians: Open-Vocabulary Scene Understanding with 3D Gaussian Splatting. arXiv preprint arXiv:2403.15624 (2024). https://doi.org/10.48550/arXiv.2403.15624

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.15624 2024

  17. [17]

    arXiv preprint arXiv:2508.08252 (2025)

    He, S., Jie, G., Wang, C., Zhou, Y., Hu, S., Li, G., Ding, H.: ReferSplat: Referring Segmentation in 3D Gaussian Splatting. arXiv preprint arXiv:2508.08252 (2025)

    Pith/arXiv arXiv 2025

  18. [18]

    Direct Determina- tion by Recursion

    Ivanic, J., Ruedenberg, K.: Rotation Matrices for Real Spherical Harmonics. Direct Determina- tion by Recursion. The Journal of Physical Chemistry100(15), 6342–6347 (1996)

    1996

  19. [19]

    NeurIPS37, 89184–89212 (2024) 16

    Jain, U., Mirzaei, A., Gilitschenski, I.: GaussianCut: Interactive Segmentation via Graph Cut for 3D Gaussian Splatting. NeurIPS37, 89184–89212 (2024) 16

    2024

  20. [20]

    ACM Transactions on Graphics42(4) (2023)

    Kerbl, B., Kopanas, G., Leimk¨ uhler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics42(4) (2023). https://doi.org/10.1145/3592433.3592553

    arXiv 2023

  21. [21]

    arXiv preprint arXiv:2510.22984 (2025)

    Kim, C., Zhao, S., Zhu, M., Lin, T.Y., Ghaffari, M.: Equivariant Neural Networks for General Linear Symmetries on Lie Algebras. arXiv preprint arXiv:2510.22984 (2025). https://doi.org/10.48550/arXiv.2510.22984

    doi:10.48550/arxiv.2510.22984 2025

  22. [22]

    arXiv preprint arXiv:2506.09565 (2025)

    Li, Q., Sun, J., An, L., Su, Z., Zhang, H., Liu, Y.: SemanticSplat: Feed-Forward 3D Scene Understanding with Language-Aware Gaussian Fields. arXiv preprint arXiv:2506.09565 (2025). https://doi.org/10.48550/arXiv.2506.09565

    Pith/arXiv arXiv doi:10.48550/arxiv.2506.09565 2025

  23. [23]

    arXiv preprint arXiv:2503.18052 (2025)

    Li, Y., Ma, Q., Yang, R., Li, H., Ma, M., Ren, B., Popovic, N., Sebe, N., Konukoglu, E., Gevers, T., Van Gool, L., Oswald, M.R., Paudel, D.P.: Scenesplat: Gaussian Splatting-based Scene Understanding with Vision-Language Pretraining. arXiv preprint arXiv:2503.18052 (2025). https://doi.org/10.48550/arXiv.2503.18052

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.18052 2025

  24. [24]

    arXiv preprint arXiv:2306.12059 (2023)

    Liao, Y.L., Wood, B., Das, A., Smidt, T.: EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations. arXiv preprint arXiv:2306.12059 (2023)

    Pith/arXiv arXiv 2023

  25. [25]

    30529–30545 (2024)

    Lin, T.Y., Zhu, M., Ghaffari, M.: Lie Neurons: Adjoint-Equivariant Neural Networks for Semisimple Lie Algebras pp. 30529–30545 (2024)

    2024

  26. [26]

    In: ECCV

    Lu, G., Zhang, S., Wang, Z., Liu, C., Lu, J., Tang, Y.: ManiGaussian: Dynamic Gaussian Splatting for Multi-task Robotic Manipulation. In: ECCV. pp. 349–366. Springer (2024)

    2024

  27. [27]

    arXiv preprint arXiv:2506.08710 (2025)

    Ma, M., Ma, Q., Li, Y., Cheng, J., Yang, R., Ren, B., Popovic, N., Wei, M., Sebe, N., Van Gool, L., Gevers, T., Oswald, M.R., Paudel, D.P.: SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting. arXiv preprint arXiv:2506.08710 (2025). https://doi.org/10.48550/arXiv.2506.08710

    doi:10.48550/arxiv.2506.08710 2025

  28. [28]

    arXiv preprint arXiv:2408.10906 (2024)

    Ma, Q., Li, Y., Ren, B., Sebe, N., Konukoglu, E., Gevers, T., Van Gool, L., Paudel, D.P.: ShapeSplat: A Large-scale Dataset of Gaussian Splats and Their Self-Supervised Pretraining. arXiv preprint arXiv:2408.10906 (2024). https://doi.org/10.48550/arXiv.2408.10906

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.10906 2024

  29. [29]

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. vol. 65, pp. 99–106. ACM New York, NY, USA (2021)

    2021

  30. [30]

    World Scientific Annual Review of Artificial Intelligence1, 2440001 (2023)

    Pang, Y., Tay, E.H.F., Yuan, L., Chen, Z.: Masked autoencoders for 3D Point Cloud Self- supervised Learning. World Scientific Annual Review of Artificial Intelligence1, 2440001 (2023)

    2023

  31. [31]

    arXiv preprint arXiv:2410.07577 (2024)

    Peng, Q., Planche, B., Gao, Z., Zheng, M., Choudhuri, A., Chen, T., Chen, C., Wu, Z.: 3D Vision-Language Gaussian Splatting. arXiv preprint arXiv:2410.07577 (2024)

    Pith/arXiv arXiv 2024

  32. [32]

    Journal of Physics A: Mathematical and Theoretical40(7), 1597–1610 (2007)

    Pinchon, D., Hoggan, P.E.: Rotation matrices for real spherical harmonics: general rotations of atomic orbitals in space-fixed axes. Journal of Physics A: Mathematical and Theoretical40(7), 1597–1610 (2007)

    2007

  33. [33]

    652–660 (2017) 17

    Qi, C.R., Su, H., Mo, K., Guibas, L.J.: PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation pp. 652–660 (2017) 17

    2017

  34. [34]

    NeurIPS30(2017)

    Qi, C.R., Yi, L., Su, H., Guibas, L.J.: PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. NeurIPS30(2017)

    2017

  35. [35]

    9323–9332 (2021)

    Satorras, V.G., Hoogeboom, E., Welling, M.: E(n) Equivariant Graph Neural Networks pp. 9323–9332 (2021)

    2021

  36. [36]

    456–472 (2024)

    Shen, Q., Yang, X., Wang, X.: FlashSplat: 2D to 3D Gaussian Splatting Segmentation Solved Optimally pp. 456–472 (2024)

    2024

  37. [37]

    5333–5343 (2024)

    Shi, J.C., Wang, M., Duan, H.B., Guan, S.H.: Language Embedded 3D Gaussians for Open- Vocabulary Scene Understanding pp. 5333–5343 (2024)

    2024

  38. [38]

    arXiv preprint arXiv:2511.18386 (2025)

    Siegel, P., Tombari, F., Pollefeys, M., Barath, D.: SegSplat: Feed-forward Gaussian Splatting and Open-Set Semantic Segmentation. arXiv preprint arXiv:2511.18386 (2025). https://doi.org/10.48550/arXiv.2511.18386

    work page doi:10.48550/arxiv.2511.18386 2025

  39. [39]

    In: Advances in Neural Information Processing Systems (2023)

    Simeon, G., de Fabritiis, G.: Tensornet: Cartesian tensor representations for efficient learning of molecular potentials. In: Advances in Neural Information Processing Systems (2023)

    2023

  40. [40]

    10208–10217 (2024)

    Szymanowicz, S., Rupprecht, C., Vedaldi, A.: Splatter Image: Ultra-Fast Single-View 3D Reconstruction pp. 10208–10217 (2024)

    2024

  41. [41]

    1–18 (2024)

    Tang, J., Chen, Z., Chen, X., Wang, T., Zeng, G., Liu, Z.: LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation pp. 1–18 (2024)

    2024

  42. [42]

    In: ICLR (2022), https://openreview.net/forum?id= zNHzqZ9wrRB

    Th¨ olke, P., De Fabritiis, G.: TorchMD-NET: Equivariant Transformers for Neural Net- work Based Molecular Potentials. In: ICLR (2022), https://openreview.net/forum?id= zNHzqZ9wrRB

    2022

  43. [43]

    arXiv preprint arXiv:1802.08219 (2018)

    Thomas, N., Smidt, T., Kearnes, S., Yang, L., Li, L., Kohlhoff, K., Riley, P.: Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds. arXiv preprint arXiv:1802.08219 (2018)

    Pith/arXiv arXiv 2018

  44. [44]

    Reviews of Modern Physics 52(2), 299–339 (1980)

    Thorne, K.S.: Multipole Expansions of Gravitational Radiation. Reviews of Modern Physics 52(2), 299–339 (1980). https://doi.org/10.1103/RevModPhys.52.299

    doi:10.1103/revmodphys.52.299 1980

  45. [45]

    World Scientific (1988)

    Varshalovich, D.A., Moskalev, A.N., Khersonskii, V.K.: Quantum Theory of Angular Momen- tum. World Scientific (1988)

    1988

  46. [46]

    241–258 (2022)

    Wad, T., Sun, Q., Pranata, S., Jayashree, K., Zhang, H.: Equivariance and Invariance Inductive Bias for Learning from Insufficient Data pp. 241–258 (2022)

    2022

  47. [47]

    The Surprising Effectiveness of Equivariant Models in Domains with Latent Symmetry

    Wang, D., Park, J.Y., Sortur, N., Wong, L.L., Walters, R., Platt, R.: The Surprising Effective- ness of Equivariant Models in Domains with Latent Symmetry. arXiv preprint arXiv:2211.09231 (2022). https://doi.org/10.48550/arXiv.2211.09231, iCLR 2023

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.09231 2022

  48. [48]

    NeurIPS31(2018)

    Weiler, M., Geiger, M., Welling, M., Boomsma, W., Cohen, T.S.: 3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data. NeurIPS31(2018)

    2018

  49. [49]

    Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point Transformer V2: Grouped Vector Attention and Partition-based Pooling (2022) 18

    2022

  50. [50]

    In: ICML (2025), https://openreview

    Xie, Y., Daigavane, A., Kotak, M., Smidt, T.: The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products. In: ICML (2025), https://openreview. net/forum?id=EvIwwGYTLc

    2025

  51. [51]

    In: ECCV

    Ye, M., Danelljan, M., Yu, F., Ke, L.: Gaussian Grouping: Segment and Edit Anything in 3D Scenes. In: ECCV. pp. 162–179. Springer (2024)

    2024

  52. [52]

    19313–19322 (2022)

    Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-BERT: Pre-training 3D Point Cloud Transformers with Masked Point Modeling pp. 19313–19322 (2022)

    2022

  53. [53]

    Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point Transformer pp. 16259–16268 (2021) 19 A Explicitgl(3)Intertwiners for SH Degreesℓ≤2 We provide an explicit realization of real spherical harmonics of degrees ℓ = 0, 1, 2 inside the 3 × 3 matrix carriergl(3). Throughout,R∈SO(3) acts ongl(3) by conjugation, Ad R(M) :=RM R⊤. Under this action, gl(3)...

    2021