Pith. sign in

REVIEW 3 major objections 5 minor 5 cited by

View-synthesis loss alone cannot teach Gaussian splats geometry; G3Splat's two differentiable priors (ray alignment, surface-normal alignment) make feed-forward splatting well-posed, yielding pose, depth, and mesh from two unposed images.

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

2026-08-03 15:13 UTC pith:KGLBDZPI

load-bearing objection Useful priors, real geometric gains, but the 3DGS ablation for the orientation loss is confounded with the scale-flattening term, so the causal story is stronger than the evidence. the 3 major comments →

arxiv 2512.17547 v2 pith:KGLBDZPI submitted 2025-12-19 cs.CV

G3Splat: Geometrically Consistent Generalizable Gaussian Splatting

Mehdi Hosseinzadeh , Shin-Fang Chng , Yi Xu , Simon Lucey , Ian Reid , Ravi Garg This is my paper
classification cs.CV
keywords 3D Gaussian splattinggeneralizable splattinggeometric priorsrelative pose estimationnovel-view depthself-supervised learningsurface normalsfeed-forward reconstruction
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.

The paper's central claim is that training feed-forward networks to predict per-pixel 3D Gaussians from images using only a view-synthesis loss is ill-posed: many different 3D Gaussian configurations can produce the same rendered views, so the learned orientations and scales drift into geometric nonsense. G3Splat shows that adding two differentiable geometric priors resolves this. The first, Lalign, forces each Gaussian center to reproject exactly to its pixel, keeping Gaussians on their viewing rays. The second, Lorient, estimates a local surface normal from the cross product of neighboring Gaussian means and supervises the Gaussian's smallest-scale eigenvector to agree with it, giving orientation direct geometric meaning. Trained on RealEstate10K, the model achieves state-of-the-art relative pose estimation, novel-view depth, and TSDF-fused mesh reconstruction on zero-shot datasets like ScanNet and ACID, while maintaining or slightly improving novel-view synthesis quality. The priors are architecture-agnostic and work with both DUSt3R-style and VGGT-style backbones.

Core claim

On its own terms, the paper diagnoses a degeneracy: when a splat predictor is trained with only photometric losses, the Gaussian means (which come from a strong depth/point backbone) remain roughly sensible, but orientations, scales, and opacities are free to absorb appearance rather than geometry. The authors then show that a pixel-alignment loss (Lalign) and an edge-aware orientation loss (Lorient) — the latter using the cross product of central-difference neighbor means as a pseudo surface normal — eliminate the degeneracy. With these, predicted Gaussians align with real surface structure, which makes it possible to render novel-view depth maps reliably, fuse them into meshes via TSDF, an

What carries the argument

The central object is the per-pixel 3D Gaussian splat — mean, quaternion orientation, three scale parameters, opacity, color — predicted in a canonical coordinate frame. The argument turns on two differentiable regularizers: Lalign, a re-projection loss that demands each Gaussian center reproject exactly to its pixel (so splats cannot drift off their viewing rays), and Lorient, which computes a local surface normal from the cross product of central-difference vectors between neighboring Gaussian means (Δy × Δx), weights it by an edge-aware confidence term, and penalizes angular disagreement with the Gaussian's smallest-eigenvalue direction. The orientation loss provides the missing structura

Load-bearing premise

The orientation prior assumes the cross product of neighboring predicted Gaussian centers is the true local surface normal, which holds only where means are accurate and the surface is smooth; at depth discontinuities, occlusions, or with mean errors, these pseudo-labels are wrong, and edge-weighting merely downweights rather than corrects them.

What would settle it

Train a generalizable splatting model with the same two priors on a dataset dominated by thin structures, vegetation, or heavy texture (where central-difference normals are unreliable) and compare against the view-synthesis-only baseline: if relative pose and depth gains vanish or reverse, the assumption that local finite-difference normals provide a trustworthy supervisory signal in new domains would be refuted. A more direct test is to inspect predicted Gaussian normals along a rendered depth discontinuity: correct geometry should show normals flipping exactly at the edge, whereas the prior'

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Any feed-forward geometry backbone (depth-based or point-map-based) can be turned into a geometrically meaningful splat predictor by adding Lalign and Lorient, so future architectures inherit the well-posedness without new supervision.
  • A single RE10K-trained model can serve as a general structure-and-pose estimator, producing relative pose, depth, and meshes from two unposed images — matching or beating task-specific matching methods on zero-shot indoor scenes.
  • Geometric evaluation of splatting methods becomes possible: rendered depth from virtual views and TSDF-fused meshes can be used as standard benchmarks, replacing input-view-only depth metrics.
  • Test-time photometric pose refinement using the same losses further improves pose and depth, indicating the priors are stable enough to guide gradient-based optimization at inference.
  • The edge-aware normal prior and flatness penalty push full-rank 3DGS splats toward a surfel-like behavior, narrowing the gap between 3DGS and 2DGS representations in the generalizable setting.

Where Pith is reading between the lines

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

  • The paper's diagnosis generalizes: any representation with more degrees of freedom than needed to explain pixels (radiance fields, warped surfaces) may exhibit the same photometric degeneracy; the alignment-plus-orientation recipe could transfer to other overparameterized self-supervised learners.
  • One testable extension is to replace the central-difference normal estimator with a robust plane fit or learned normal predictor; if that reduces the need for the edge-aware weighting, the orientation prior could become stronger in geometrically complex scenes.
  • Since the priors only supervise orientation through the predicted means, any systematic bias in the mean predictor (e.g., scale drift) would propagate to normals; coupling Lorient to a scale-invariant normal loss could decouple mean accuracy from orientation accuracy.
  • If the priors truly make the problem well-posed, they may also stabilize few-shot per-scene Gaussian optimization, offering a replacement for depth-regularization heuristics currently used in sparse-view 3DGS.

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

3 major / 5 minor

Summary. The paper proposes G3Splat, a framework for generalizable Gaussian splatting that augments the standard view-synthesis loss with two geometric priors: a pixel-alignment loss Lalign and an orientation-alignment loss Lorient, plus a minimum-scale penalty Lflat for 3DGS. The method is instantiated on DUSt3R- and VGGT-based backbones with both 3DGS and 2DGS representations, trained on RealEstate10K, and evaluated on RE10K, ACID, and ScanNet for relative pose, depth, mesh, and novel-view synthesis. The central claim is that view-synthesis supervision alone is under-constrained, and the proposed priors make predicted Gaussian orientations and scales geometrically meaningful.

Significance. If the central claim holds, the paper addresses a genuine degeneracy in generalizable Gaussian splatting and provides a simple, architecture-agnostic remedy. Strengths include extensive evaluation with retrained NoPoSplat baselines, consistent ablations across two backbones and two scene representations, a well-motivated expected-depth renderer, and released code and models. The main concern is that the ablation evidence for the orientation prior is confounded with Lflat for the primary 3DGS variant, and the mesh evaluation omits external baselines despite SOTA claims; both weaken the paper's central attribution as currently presented.

major comments (3)
  1. [§3.1, Eq. (12); Tables 2, 4, 9, 10] For the 3DGS variant, Lflat is applied 'whenever Lorient is enabled' and all ablation rows labeled 'Orient' therefore include both Lorient and Lflat. No row isolates Lflat alone or Lorient alone. Because Lflat directly penalizes the minimum Gaussian scale, it could by itself reduce geometric degeneracy and improve pose/depth, so the reported gains of 'Orient' over 'Align' cannot be attributed to orientation alignment. This is load-bearing: the paper's headline conceptual finding—that aligning orientations with surface normals is crucial—is argued primarily from the 3DGS results. The 2DGS results, where Lflat is disabled, do not cleanly support it either: '2DGS+Orient' alone is unstable (e.g., Table 2, RE10K Trans. 34.414 vs. 29.693 no-prior). Please add ablations that separate Lflat and Lorient (e.g., Align+Lflat, Align+Orient without Lflat) before claiming orientation alignment is the d
  2. [§4.2, Table 5] The mesh reconstruction table compares only internal variants (No Prior/With Priors for DUSt3R and VGGT). The abstract and Section 4.2 claim state-of-the-art mesh reconstruction, but no external prior-art method (e.g., MVSplat, NoPoSplat, DepthSplat) is included in the mesh comparison, despite such baselines appearing in other tables. Either include external methods under the same TSDF-Fusion protocol, or revise the claims to state that the comparison is an internal ablation. This is necessary to support the 'state-of-the-art mesh reconstruction' assertion.
  3. [§4.2, Tables 3 and 4] Depth evaluation uses ground-truth relative poses for rendering virtual views, which the text notes places pose-free methods at a disadvantage. While this is a defensible protocol, the reported margin over NoPoSplat (AbsRel 0.090 vs 0.131) is large and central to the geometry claim. However, because the 3DGS ablation columns again combine Lorient with Lflat, the depth gains in Table 4 inherit the same confound as the pose ablations. If Lflat is responsible for most of the improvement, the depth and mesh conclusions would need to be reframed accordingly.
minor comments (5)
  1. [General] No error bars, confidence intervals, or multiple-seed runs are reported. Several headline differences are small (e.g., Table 7 average PSNR 23.417 vs. 23.244 for NoPoSplat; Table 1 ScanNet AUC@5 0.124 vs. 0.109 for NoPoSplat w/ refinement). Reporting variance across seeds would strengthen the quantitative claims.
  2. [§4.3] The sentence 'We noted that the proposed method outperforms prior work...' appears without a reference to the underlying table; please cite Table 7 (and Table 8) explicitly.
  3. [Eq. (9)] The notation with trailing asterisk for normalization is nonstandard; define it where first used or use a named operator.
  4. [§3.1, Eq. (11)] The definition of Ω says it excludes a one-pixel boundary, but |Ω|=T(H−2)(W−2) assumes all interior pixels are valid; in practice invalid depths/out-of-bounds means may further reduce the set. Clarify whether Eq. (11) normalizes by the count of valid pixels or by the full interior size.
  5. [Introduction / Related Work] The statement that 'all existing generalizable splatting methods struggle to learn geometrically faithful orientations and scales' is strong; consider softening or citing the specific evidence for each cited method.

Circularity Check

0 steps flagged

No significant circularity: the claimed geometric gains are measured against external ground truth, and the priors are self-consistency regularizers rather than fitted inputs renamed as predictions.

full rationale

The paper's derivation chain is self-contained with respect to circularity. The orientation prior Lorient (Eqs. 8-11) supervises predicted Gaussian normals using normals computed by finite differences of predicted Gaussian means; this is self-consistency between two predicted outputs, not a fitted quantity masquerading as an external prediction. The pixel-alignment loss Lalign (Eqs. 13-14) uses training-time ground-truth intrinsics and extrinsics as a regularizer, but test-time pose/depth/mesh evaluations are performed against held-out COLMAP poses, ScanNet RGB-D ground truth, and TSDF meshes, none of which appear as training targets in the priors. The scale regularizer Lflat (Eq. 12) is a genuine ablation confound for the 3DGS variant: the paper states it is applied 'whenever Lorient is enabled' (Sec. 3.1), so the 'Orient' rows in Tables 2, 4, 9, and 10 do not isolate orientation alignment from flattening. This is a correctness/attribution concern, not a circularity: the reported metrics are not constructed from the losses, and the 2DGS experiments, where Lflat is disabled, provide partially supporting evidence for the orientation prior. Self-citations appear only in related-work and implementation contexts and are not load-bearing for the central claim. No step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method is empirical; no analytical derivation of well-posedness is given. The central geometric idea relies on two assumptions: local finite-difference normals are meaningful surface pseudo-labels, and training-time camera poses can be used to regularize pixel alignment without undermining pose-free generalization. The listed hyperparameters are hand-chosen and fixed in §8.

free parameters (7)
  • lambda_a (alignment weight) = 0.1
    Hand-chosen loss weight for Lalign; fixed across all experiments (§8).
  • lambda_o (orientation weight) = 0.05
    Hand-chosen loss weight for Lorient; fixed across all experiments (§8).
  • lambda_flat (minimum-scale penalty weight) = 1000
    Hand-chosen penalty on the minimum Gaussian scale for 3DGS only (§3.1, Eq. 12).
  • w0 (edge-aware weight constant) = 10
    Fixed constant in edge-aware weight formula Eq. (10), §8.
  • kappa (edge decay) = 4
    Fixed constant in edge-aware weight formula Eq. (10), §8.
  • q (quantile for normalization) = 0.95
    Quantile used to normalize local 3D variation in Eq. (10), §8.
  • delta (Huber threshold) = 0.1
    Threshold for the cosine-space Huber penalty in Eq. (11), §8.
axioms (5)
  • standard math Cross product of central-difference vectors (Δy × Δx) gives the local surface normal.
    Invoked in Eq. (9); exact only when the three neighboring points are coplanar in the local tangent plane; approximation under noise and non-planar geometry.
  • domain assumption Gaussian centers predicted for neighboring pixels in each input frame are spatially adjacent on the underlying surface.
    Needed for central-difference normals in §3.1; breaks at depth discontinuities, occlusions, and foreground/background boundaries; edge-weighting mitigates but does not correct.
  • domain assumption COLMAP or ground-truth poses used by Lalign during training are accurate enough to define pixel-alignment targets.
    Eqs. (13)-(14) use Kt, Rt, Tt; if training poses are wrong, the alignment prior injects error, and pose-free inference relies on this train-time supervision.
  • domain assumption A Gaussian's smallest-eigenvalue eigenvector is the geometric normal of the underlying surface.
    Uses Eq. (7) and the definition of Gaussian normal; for full-rank 3DGS this holds only if the splat is flat, which is encouraged by Lflat—a bias that can be wrong for volumetric regions.
  • domain assumption Pretrained geometry backbones (MASt3R/VGGT) provide a sufficient starting point for fine-tuning with view-synthesis and the proposed priors.
    Both variants initialize from supervised pretrained models (§8); the reported geometric gains partly inherit this supervision, so the 'self-supervised' label is only true after pretraining.

pith-pipeline@v1.3.0-alltime-deepseek · 29077 in / 16757 out tokens · 172887 ms · 2026-08-03T15:13:16.041515+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of G3Splat: Geometrically Consistent Generalizable Gaussian Splatting." pith.science (2026). https://pith.science/paper/KGLBDZPI

@misc{pith2026251217547,
  author       = {Pith},
  title        = {Pith review of: G3Splat: Geometrically Consistent Generalizable Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGLBDZPI}},
  note         = {Machine review of arXiv:2512.17547}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussians have become a powerful scene representation for real-time splatting and high-quality novel-view synthesis. This has motivated generalizable splatting -- methods that adapt feed-forward geometry prediction networks to produce per-pixel Gaussians from a set of images. However, most generalizable splatting pipelines are supervised primarily through a view-synthesis loss to predict Gaussian orientation, anisotropic scale, opacity, and appearance in addition to their locations. We show that this learning objective is under-constrained. Models trained with view synthesis alone produce splats whose orientations and scales have no geometric connotation. The result is that, while producing decent view-synthesis performance, nearly all generalizable splatting methods produce geometrically inaccurate and misaligned Gaussians. We introduce G3Splat, a geometry-consistent generalizable splatting framework that addresses these degeneracies through differentiable geometric priors on the predicted 3D Gaussians, making the learning problem well-posed. These priors encourage the per-pixel splats to remain on their viewing rays and to orient themselves in accordance with local surfaces. Our priors are architecture-agnostic and can be incorporated into any previously studied geometric backbone for generalizable splatting, as well as different scene representations. We test G3Splat with both DUSt3R-style and VGGT-style backbones to predict pixel-aligned full-rank 3DGS as well as surfel-like 2DGS. Trained on RE10K, G3Splat produces Gaussian splats with significantly higher geometric fidelity than baselines, providing state-of-the-art novel-view depth, mesh reconstruction, and relative pose estimation performance while preserving novel-view synthesis quality, as evaluated on datasets such as ACID and ScanNet. Code and pretrained models are released on our project page.

Figures

Figures reproduced from arXiv: 2512.17547 by Ian Reid, Mehdi Hosseinzadeh, Ravi Garg, Shin-Fang Chng, Simon Lucey, Yi Xu.

Figure 1
Figure 1. Figure 1: G 3 Splat enables geometrically consistent, pose-free generalizable Gaussian splatting across backbones. Left: our VGGT￾based [56] adaptation without / with the proposed priors. Right: our DUSt3R-based [58] adaptation without / with the proposed priors. We visualize reconstructions on a Sora-generated video (150 input views) and RealEstate10K [75] (2 input views). Our priors encourage geometrically consist… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of predicted Gaussian parameters. For visualization, we denote by (s1, s2, s3) the sorted eigen-scales of each Gaussian covariance in Equation (7), such that s1 ≥ s2 ≥ s3; the smallest scale s3 characterizes uncertainty along the surface normal direction. Row 1 (ours) shows: (a) the source image to which Gaussians are aligned, (b) skewness of the estimated Gaussians within their defi… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of rendered novel-view depth on RE10K (first row), ACID (second row), and ScanNet (last row). We report average translation and rotation errors (in de￾grees), where the (scale-invariant) translation error is defined as the angular difference between normalized predicted and ground-truth translation vectors. The results show that our priors yield significantly lower translation and ro… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative ablation of reconstructed meshes on ScanNet [9] (2 input views) using VGGT [56] and DUSt3R [58] backbones. Our proposed priors consistently yield sharper, more complete, and less noisy geometry across both backbones [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative ablation of reconstructed Gaussians on a Sora-generated video (VGGT backbone, 24 input views). Prompt used to generate the video: “A single unbroken orbital camera move through a vast, empty gothic library, with static architecture, medium-wide framing, warm steady lighting, and crisp sharp geometric details”. tors: MonoDepthV2 [20], SC-SfM-Learners [1], and SC￾DepthV3 [49]. On this NYUD-v2 [47… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative ablation of reconstructed Gaussians on RE10K [75] (VGGT backbone, 24 input views). target pose relative to the first input image. In contrast, meth￾ods that predict Gaussian means as depth maps require an additional warping step, using the ground-truth relative pose, to align Gaussians to the first-view coordinate frame. Al￾though some works (e.g., [48]) report novel-view results for DUSt3R [58… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative ablation of reconstructed Gaussians on Tanks and Temples [32] (VGGT backbone, 20 input views) [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative ablation of reconstructed Gaussians on RE10K [75] (DUSt3R backbone, 2 input views) [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative ablation of the losses. We visualize the learned Gaussian orientations (from the first input view) along with rendered novel-view depth, color, and surface normals, on RE10K [75]. As shown, using Lalign alone is insufficient to learn reliable Gaussian orientations; adding Lorient encourages geometry-consistent orientations and, together, they produce more accurate rendered depth. the pose-free,… view at source ↗
Figure 10
Figure 10. Figure 10: More qualitative comparison of novel-view rendered depth on RE10K [75] and ScanNet [9]. pixelSplat depths are relatively geometrically consistent but noisy when the baseline is small. Large errors can be observed in the pixelSplat depths when the image overlap is small. Other baselines provide depth maps, which are hypersensitive to image texture. While some potentially meaningful fine structural edges ar… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of mesh reconstruction on ScanNet [9] (DUSt3R backbone, 2 input views). For each scene, we show the two input context views, the textured mesh reconstructed by fusing virtual rendered depth maps via TSDF-Fusion [8], the Gaussian normals for the first input view, and the ground-truth and rendered depth from a novel (virtual) viewpoint. Baselines exhibit inaccurate rendered depth and … view at source ↗
Figure 12
Figure 12. Figure 12: Failure modes of rendered normal–depth consistency compared to our orientation loss. We replace the proposed orientation prior Lorient with the rendered normal–depth consistency loss LRNC [26] and visualize the learned scene during training on RE10K [75], by projecting the reconstructed 3D Gaussians onto three axis-aligned planes. Using LRNC without the alignment loss Lalign yields severely degenerate rec… view at source ↗

discussion (0)

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

Forward citations

Cited by 5 Pith papers

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

  1. DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis

    cs.CV 2026-04 unverdicted novelty 8.0

    DF3DV-1K supplies 1,048 scenes with clean and cluttered image pairs plus a challenging 41-scene subset to benchmark and improve distractor-free radiance field methods.

  2. G$^2$SR: Geometric Methods for Fast and Memory-Efficient Gaussian-based Surface Reconstruction

    cs.CV 2026-07 conditional novelty 7.0

    A lightweight neural frontend plus analytic triangulation recovers metric-scale 3D Gaussian surfaces from 2–3 posed images faster and in less memory than end-to-end splat regressors.

  3. DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis

    cs.CV 2026-04 unverdicted novelty 6.0

    DF3DV-1K supplies 1,048 real scenes with clean/cluttered image pairs and a 41-scene hard subset to benchmark and improve distractor-free radiance-field methods.

  4. RMGS-SLAM: Real-time Multi-sensor Gaussian Splatting SLAM

    cs.RO 2026-04 unverdicted novelty 6.0

    RMGS-SLAM delivers real-time multi-sensor 3D Gaussian splatting SLAM with cascaded initialization, Gaussian-based loop closure, and claimed state-of-the-art efficiency, accuracy, and rendering quality on large-scale l...

  5. FlowObject: Flow Steering for Bridging Generative Priors and Reconstruction Fidelity

    cs.CV 2026-06 unverdicted novelty 4.0

    FlowObject reformulates sparse-view 3D reconstruction as a training-free guided inverse problem in flow-matching models, augmented by 3DGS refinement to improve geometric completeness and fidelity.

Reference graph

Works this paper leans on

79 extracted references · 5 linked inside Pith · cited by 4 Pith papers

  1. [1]

    Auto-rectify network for un- supervised indoor depth estimation.IEEE transactions on pattern analysis and machine intelligence, 44(12):9802–9813,

    Jia-Wang Bian, Huangying Zhan, Naiyan Wang, Tat-Jun Chin, Chunhua Shen, and Ian Reid. Auto-rectify network for un- supervised indoor depth estimation.IEEE transactions on pattern analysis and machine intelligence, 44(12):9802–9813,

  2. [2]

    Pyramid stereo match- ing network

    Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo match- ing network. InCVPR, 2018. 3

    2018

  3. [3]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. InCVPR,

  4. [4]

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

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction.TVCG,

  5. [5]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. InECCV, 2024. 1, 2, 3, 4, 6, 8, 5

    2024

  6. [6]

    Invertible neural warp for nerf

    Shin-Fang Chng, Ravi Garg, Hemanth Saratchandran, and Simon Lucey. Invertible neural warp for nerf. InECCV, 2024. 3

    2024

  7. [7]

    Depth- regularized optimization for 3d gaussian splatting in few-shot images

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth- regularized optimization for 3d gaussian splatting in few-shot images. InCVPR, 2024. 2, 3

    2024

  8. [8]

    A volumetric method for building complex models from range images

    Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. InProceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303–312, 1996. 9, 5, 12, 13

    1996

  9. [9]

    Scannet: Richly- annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly- annotated 3d reconstructions of indoor scenes. InCVPR,

  10. [10]

    Learning to render novel views from wide-baseline stereo pairs

    Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitz- mann. Learning to render novel views from wide-baseline stereo pairs. InCVPR, 2023. 5

    2023

  11. [11]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. Roma: Robust dense feature matching. InCVPR, 2024. 2, 6, 7

    2024

  12. [12]

    Predicting depth, surface nor- mals and semantic labels with a common multi-scale convolu- tional architecture

    David Eigen and Rob Fergus. Predicting depth, surface nor- mals and semantic labels with a common multi-scale convolu- tional architecture. InProceedings of the IEEE international conference on computer vision, pages 2650–2658, 2015. 5

    2015

  13. [13]

    Depth map prediction from a single image using a multi-scale deep network

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. InNeurIPS, 2014. 3

    2014

  14. [14]

    Deep ordinal regression network for monocular depth estimation

    Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat- manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. InCVPR, 2018. 3

    2018

  15. [15]

    Colmap-free 3d gaussian splatting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. InCVPR, 2024. 4

    2024

  16. [16]

    Unsupervised cnn for single view depth estimation: Geometry to the rescue

    Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. InECCV, 2016. 4

    2016

  17. [17]

    Direct align- ment for robust nerf learning

    Ravi Garg, Shin-Fang Chng, and Simon Lucey. Direct align- ment for robust nerf learning. InACCV, 2024. 3

    2024

  18. [18]

    Unsupervised monocular depth estimation with left-right con- sistency

    Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right con- sistency. InCVPR, 2017. 4

    2017

  19. [19]

    Clement Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J. Brostow. Digging into self-supervised monocular depth estimation. InICCV, 2019. 4

    2019

  20. [20]

    Digging into self-supervised monocular depth estimation

    Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 3828–3838,

  21. [21]

    Cascade cost volume for high-resolution multi-view stereo and stereo matching

    Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. InCVPR, 2020. 3

    2020

  22. [22]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering.CVPR, 2024

    Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering.CVPR, 2024. 3

    2024

  23. [23]

    Neural 3d scene reconstruction with the manhattan-world assumption

    Haoyu Guo, Sida Peng, Haotong Lin, Qianqian Wang, Guofeng Zhang, Hujun Bao, and Xiaowei Zhou. Neural 3d scene reconstruction with the manhattan-world assumption. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5511–5520, 2022. 9, 14

    2022

  24. [24]

    Binocular-guided 3d gaussian splatting with view consistency for sparse view synthesis.NeurIPS, 2024

    Liang Han, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Binocular-guided 3d gaussian splatting with view consistency for sparse view synthesis.NeurIPS, 2024. 3

    2024

  25. [25]

    Unifying corre- spondence pose and nerf for generalized pose-free novel view synthesis

    Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying corre- spondence pose and nerf for generalized pose-free novel view synthesis. InCVPR, 2024. 6, 7, 5

    2024

  26. [26]

    2d gaussian splatting for geometrically accu- rate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accu- rate radiance fields. InACM SIGGRAPH 2024, 2024. 2, 3, 4, 5, 10, 11, 13

    2024

  27. [27]

    Deepmvs: Learning multi-view stereopsis

    Jia-Bin Huang, Iain Matthews, and Wolf Kienzle. Deepmvs: Learning multi-view stereopsis. InCVPR, 2018. 3

    2018

  28. [28]

    Anysplat: Feed-forward 3d gaussian splatting from unconstrained views.arXiv preprint arXiv:2505.23716,

    Lihan Jiang, Yucheng Mao, Linning Xu, Tao Lu, Kerui Ren, Yichen Jin, Xudong Xu, Mulin Yu, Jiangmiao Pang, Feng Zhao, et al. Anysplat: Feed-forward 3d gaussian splatting from unconstrained views.arXiv preprint arXiv:2505.23716,

    arXiv

  29. [29]

    Self- splat: Pose-free and 3d-prior-free generalizable 3d gaussian splatting.CVPR, 2025

    Seunghyun Kang, Hyunwoo Lee, and Hyeongju Chae. Self- splat: Pose-free and 3d-prior-free generalizable 3d gaussian splatting.CVPR, 2025. 1, 4, 7, 8, 5

    2025

  30. [30]

    Splatam: Splat, track & map 3d gaussians for dense rgb-d slam

    Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat, track & map 3d gaussians for dense rgb-d slam. InCVPR, 2024. 3

    2024

  31. [31]

    3d gaussian splatting for real-time radiance field rendering.ACM TOG, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM TOG, 2023. 1, 2, 3, 10

    2023

  32. [32]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM TOG, 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM TOG, 2017. 4 15

    2017

  33. [33]

    Deeper depth prediction with fully convolutional residual networks

    Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Fed- erico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In3DV, 2016. 3

    2016

  34. [34]

    Bts: Depth estimation via local planar guidance.arXiv preprint arXiv:1907.10326, 2019

    Jin Han Lee, Youngbok Bae, and In So Kweon Han. Bts: Depth estimation via local planar guidance.arXiv preprint arXiv:1907.10326, 2019. 3

    Pith/arXiv arXiv 1907

  35. [35]

    Ground- ing image matching in 3d with mast3r.arXiv preprint arXiv:2406.09756, 2024

    Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Ground- ing image matching in 3d with mast3r.arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 4, 6, 7, 10

    Pith/arXiv arXiv 2024

  36. [36]

    Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. InCVPR, 2024. 3

    2024

  37. [37]

    Sgs-slam: Se- mantic gaussian splatting for neural dense slam

    Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. Sgs-slam: Se- mantic gaussian splatting for neural dense slam. InECCV,

  38. [38]

    Megadepth: Learning single- view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single- view depth prediction from internet photos. InCVPR, 2018. 6

    2018

  39. [39]

    Spacetime gaus- sian feature splatting for real-time dynamic view synthesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaus- sian feature splatting for real-time dynamic view synthesis. InCVPR, 2024. 3

    2024

  40. [40]

    Infinite nature: Perpetual view generation of natural scenes from a single image

    Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Maka- dia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. InICCV, 2021. 6, 7, 4, 5, 11, 13

    2021

  41. [41]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. InCVPR, 2024. 3, 6

    2024

  42. [42]

    Occupancy networks: Learning 3d reconstruction in function space

    Lars Mescheder, Michael Oechsle, Michael Niemeyer, Se- bastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460–4470, 2019. 14

    2019

  43. [43]

    Epipolar-free 3d gaussian splatting for generalizable novel view synthesis.NeurIPS, 2024

    Zhiyuan Min, Yawei Luo, Jianwen Sun, and Yi Yang. Epipolar-free 3d gaussian splatting for generalizable novel view synthesis.NeurIPS, 2024. 4

    2024

  44. [44]

    Creating video from text, 2024

    OpenAI. Creating video from text, 2024. 1

    2024

  45. [45]

    Vi- sion transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InICCV, 2021. 3

    2021

  46. [46]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InCVPR, 2016. 6

    2016

  47. [47]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. InComputer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7- 13, 2012, Proceedings, Part V 12, pages 746–760. Springer,

    2012

  48. [48]

    Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs.arXiv preprint arXiv:2408.13912, 2024

    Brandon Smart, Chuanxia Zheng, Iro Laina, and Vic- tor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs.arXiv preprint arXiv:2408.13912, 2024. 1, 4, 7, 8, 3, 5

    Pith/arXiv arXiv 2024

  49. [49]

    Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes.IEEE trans- actions on pattern analysis and machine intelligence, 46(1): 497–508, 2023

    Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes.IEEE trans- actions on pattern analysis and machine intelligence, 46(1): 497–508, 2023. 2, 4

    2023

  50. [50]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. InCVPR,

  51. [51]

    Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image

    Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. In2025 Interna- tional Conference on 3D Vision (3DV), pages 670–681. IEEE,

  52. [52]

    Hisplat: Hierarchical 3d gaus- sian splatting for generalizable sparse-view reconstruction

    Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaus- sian splatting for generalizable sparse-view reconstruction. ICLR, 2025. 4

    2025

  53. [53]

    Demon: Depth and motion network for learning monocular stereo

    Benjamin Ummenhofer, Hao Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. InCVPR, 2017. 3

    2017

  54. [54]

    S2gaussian: Sparse-view super-resolution 3d gaussian splatting

    Ziyu Wan, Hao Gao, Rui Xiong, and Fang Du. S2gaussian: Sparse-view super-resolution 3d gaussian splatting. InCVPR,

  55. [55]

    Patchmatchnet: Learned multi- view patchmatch stereo

    Fangjinhua Wang, Silvano Galliani, Christoph V ogel, Pablo Speciale, and Marc Pollefeys. Patchmatchnet: Learned multi- view patchmatch stereo. InCVPR, 2021. 3

    2021

  56. [56]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1, 2, 3, 8, 9, 5, 10

    2025

  57. [57]

    Efros, and Angjoo Kanazawa

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InCVPR, 2025. 3

    2025

  58. [58]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InCVPR, 2024. 1, 2, 3, 6, 7, 8, 9, 4, 5

    2024

  59. [59]

    Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes.NeurIPS, 2024

    Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes.NeurIPS, 2024. 1, 4, 8, 5, 10

    2024

  60. [60]

    4d gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. InCVPR, 2024. 3

    2024

  61. [61]

    Sparsegs: Real- time 360° sparse view synthesis using gaussian splatting

    Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real- time 360° sparse view synthesis using gaussian splatting. In 3DV, 2025. 3

    2025

  62. [62]

    Unifying flow, stereo and depth estimation.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 45(11):13941–13958,

    Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 45(11):13941–13958,

  63. [63]

    Depth- splat: Connecting gaussian splatting and depth

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depth- splat: Connecting gaussian splatting and depth. InCVPR,

  64. [64]

    Depth anything v2.arXiv:2406.09414, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2.arXiv:2406.09414, 2024. 2

    Pith/arXiv arXiv 2024

  65. [65]

    Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high- fidelity monocular dynamic scene reconstruction. InCVPR,

  66. [66]

    Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting

    Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. InICLR, 2024. 3

    2024

  67. [67]

    Mvsnet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. InECCV, 2018. 3

    2018

  68. [68]

    No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images

    Botao Ye, Sifei Liu, Haofei Xu, Li Xueting, Marc Pollefeys, Ming-Hsuan Yang, and Peng Songyou. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. InICLR, 2025. 1, 2, 3, 4, 6, 7, 8, 5, 10, 12

    2025

  69. [69]

    Scannet++: A high-fidelity dataset of 3d indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InICCV, 2023. 6

    2023

  70. [70]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 5

    2021

  71. [71]

    Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction

    Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with deep feature reconstruction. InCVPR, 2018. 4

    2018

  72. [72]

    Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis

    Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. InCVPR, 2024. 1, 4

    2024

  73. [73]

    Open3d: A modern library for 3d data processing.arXiv preprint arXiv:1801.09847, 2018

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing.arXiv preprint arXiv:1801.09847, 2018. 2

    Pith/arXiv arXiv 2018

  74. [74]

    Unsupervised learning of depth and ego-motion from video

    Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. InCVPR, 2017. 4

    2017

  75. [75]

    Stereo magnification: learning view syn- thesis using multiplane images.ACM TOG, 2018

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view syn- thesis using multiplane images.ACM TOG, 2018. 1, 6, 7, 3, 4, 5, 8, 10, 11, 13

    2018

  76. [76]

    Nice-slam: Neural implicit scalable encoding for slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 4, 9

    2022

  77. [77]

    Fsgs: Real-time few-shot view synthesis using gaussian splat- ting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splat- ting. InECCV, 2024. 2, 3

    2024

  78. [78]

    Nicer- slam: Neural implicit scene encoding for rgb slam

    Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer- slam: Neural implicit scene encoding for rgb slam. In3DV,

  79. [2017]

    2, 6, 7, 8, 9, 1, 4, 5, 10, 11, 13