Pith. sign in

REVIEW 4 major objections 4 minor 77 references

LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos

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

Pith's one-line read LongSplat claims that jointly optimizing camera poses and octree-anchored 3D Gaussians, seeded by MASt3R priors, makes novel view synthesis from casual long videos robust where drift, bad geometry initialization, and memory limits break pri

desk verdict Strong NVS results on the hardest long-video benchmarks, but the pose-SOTA claim is overstated and the COLMAP-as-ground-truth evaluation needs fixing. read the letter →

arxiv 2508.14041 v1 pith:7JEEMJZ3 submitted 2025-08-19 cs.CV

classification cs.CV
keywords novelviewsynthesis3DGaussiansplattingunposedreconstructioncameraposeestimationincrementaljointoptimizationoctreeanchorscasuallongvideoslearnedpriors
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 casually shot, camera-pose-free video of hundreds to thousands of frames can be reconstructed into a view-synthesis-ready 3D scene by a single incremental pipeline that optimizes poses and 3D Gaussians together, instead of treating pose estimation as a separate front-end. The motivation is practical: smartphone and action-camera footage regularly defeats structure-from-motion preprocessing, and existing unposed alternatives either drift, fragment, run out of memory, or inherit the noisy poses of foundation-model priors. The recipe is to seed geometry with the learned 3D prior MASt3R, compress its dense point clouds into density-adaptive octree anchors, then alternate local and global joint optimization while tracking each new frame's pose by PnP plus photometric refinement against the evolving Gaussian scene. On three benchmarks (Tanks and Temples, Free, Hike) it reports substantially higher rendering quality and lower pose error than prior unposed methods, together with roughly 8x memory compression and about 280 FPS rendering, so the claim, if true, makes long casual videos a tractable input for 3D reconstruction with no pose preprocessing at all.

What carries the argument

The load-bearing mechanism is the octree-anchored 3D Gaussian representation coupled with incremental joint optimization. Where Scaffold-GS places anchors on a fixed-resolution voxel grid seeded by SfM points, LongSplat builds anchors directly from MASt3R's dense point clouds through density-guided octree subdivision: voxels above a split threshold divide into eight children, voxels below a prune threshold are removed, and each anchor's spatial scale tracks its voxel size, so sparse regions get coarse anchors and detailed regions get fine ones. Pose estimation rides on the same priors: 2D-3D correspondences from MASt3R are solved by PnP with RANSAC, refined photometrically against the render

What would settle it

Take a long handheld sequence with independent ground-truth trajectories, for instance from a LiDAR-equipped phone, that contains long featureless or reflective stretches; run LongSplat and compare per-frame RPE and ATE. If pose error jumps sharply exactly where MASt3R's depth or correspondences are provably wrong, the claim that the joint optimization corrects the prior's failures is refuted. A cheaper check: re-render after rescaling MASt3R's depth to LiDAR depth; if quality gains vanish with the corrected prior, the prior, not the pipeline, is doing the work.

Watch

Extended reading notes

Core claim

LongSplat's central claim is that unposed novel view synthesis for casual long videos can be made robust by coupling pose estimation and scene reconstruction into one incremental process, rather than relying on a separate SfM or foundation-model front-end. The pipeline opens with MASt3R's global alignment to obtain initial poses and per-frame dense point clouds, which it converts into octree anchors whose voxel size adapts to local point density (splitting where density is high, pruning where low), an SfM-free initialization that removes redundant anchors and keeps the model around 101 MB on the Free dataset. Each new frame's pose is first solved by PnP on MASt3R correspondences back-project

Load-bearing premise

MASt3R's per-frame dense point clouds, aligned depth, and 2D correspondences must be reliable enough to anchor initialization, depth-scale alignment, anchor unprojection, and the depth supervision loss; where that learned prior fails, in textureless, occluded, or specular regions, the errors flow directly into the octree anchors and the pose estimates the joint optimization is meant to fix.

Editorial extensions

If this is right

  • If the reported numbers hold, unposed novel view synthesis becomes practical for consumer video: no SfM preprocessing, no pose front-end, and no out-of-memory failures on sequences of hundreds to thousands of frames.
  • The octree anchor design ties memory to observed geometry rather than frame count, so per-scene model size stays roughly constant as videos grow instead of exploding with resolution or duration.
  • Pose accuracy at ATE 0.004 and RPEt 0.028 on Free means the estimated trajectories are precise enough to be usable as camera paths themselves, for applications like stabilization or video-to-3D conversion, not merely as hidden scaffolding.
  • Rendering at 281.71 FPS after about one hour of training on a single RTX 4090 makes the output immediately interactively usable, a regime prior unposed long-video methods (LocalRF at 1.17 FPS) could not reach.
  • Depth plus reprojection supervision keeps newly revealed regions, where multi-view coverage is thin, well-constrained, which is likely why the method does not fragment on the Hike sequences where LocalRF visibly degrades.

Reading between the lines

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

  • Editorial inference: the visibility-IoU window selection effectively performs keyframe selection inside the optimizer, so it could be extracted as a standalone covisibility-based frame-scheduling rule for other incremental reconstruction or mapping pipelines.
  • Editorial inference: because the whole system leans on MASt3R for depth, correspondences, and the intrinsics fallback, the gains should transfer to other learned matching backbones; swapping in a metric-depth or LiDAR-depth source would directly test whether the octree plus joint-optimization machinery, rather than the prior itself, drives the improvement.
  • Editorial inference: the stated static-scene and fixed-intrinsics limitations suggest the natural next stress test is a zooming or dynamic handheld video, where the per-frame scale-aligning depth ratio and the anchor-update rule would face focal-length changes and motion breaks.
  • Editorial inference: on sequences where COLMAP itself fails, pose accuracy can only be checked against independent ground truth such as LiDAR-tracked phones; such a test would confirm the trajectory claims precisely where the paper's reference poses are known to be noisy.
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 / 4 minor

Summary. The paper proposes LongSplat, an unposed 3D Gaussian Splatting framework for novel view synthesis and camera pose estimation from casually captured long videos. The method proceeds incrementally: MASt3R provides per-frame dense point clouds, aligned depth, and 2D correspondences; octree anchors are formed from the point clouds; poses for new frames are initialized by PnP+RANSAC and refined photometrically; then local and global optimization jointly refine poses and Gaussians using photometric, depth, and reprojection losses. Experiments on Tanks and Temples, Free, and Hike report strong rendering quality, pose accuracy, training time, and memory usage compared to several baselines.

Significance. If the claims hold, LongSplat is a useful step toward practical unposed NVS on long casual videos: it combines learned 3D priors with explicit pose optimization and an adaptive anchor representation, and the reported rendering and efficiency numbers are substantially better than several prior unposed methods on the Free and Hike datasets. The per-scene tables and ablations are valuable, and the incremental formulation is clearly described. However, the paper's headline claim of 'consistently outperforms' and its pose-accuracy SOTA claim are not fully supported by the reported comparisons, so the significance is conditional on those points being resolved.

major comments (4)
  1. [§4.2 and Supplementary Table 11] HT-3DGS is absent from all main tables, yet Supp. Table 11 shows that on Tanks and Temples HT-3DGS achieves PSNR 33.53, SSIM 0.96, LPIPS 0.07, beating LongSplat's 32.83/0.94/0.08. This directly contradicts the abstract and conclusion claim of 'consistently outperforms' and 'state-of-the-art'. HT-3DGS must be included in the main comparison tables and the claim qualified according to regime (e.g., long casual videos where HT-3DGS OOMs).
  2. [§4.1, App. B.3, Tables 10 and 13] Pose accuracy on Hike is evaluated only on 6 'stable' sequences, while the hardest sequences (forest3, garden2, playground) are excluded because COLMAP fails; Table 13 shows COLMAP produces no reconstruction for these scenes. Using COLMAP poses as ground truth in precisely the regime where COLMAP is unreliable leaves the pose-accuracy SOTA claim unverified in the target setting. On Free, no independent pose reference is provided, and ATE differences (0.004 vs. 0.008) are near the scale of COLMAP noise. The paper should either provide an independent pose source or clearly limit the pose claim.
  3. [§3.2–3.3, Eqs. (8)–(13)] MASt3R supplies the initial point cloud (Sec. 3.2), the aligned depth used for scale alignment (Eq. 10), the depth supervision (Eq. 13), and the unprojected anchor points (Eq. 11). The method is therefore not independent of MASt3R's accuracy; a failure of MASt3R depth or correspondences propagates directly into the anchors and the depth loss. The paper should include a sensitivity analysis (e.g., corrupting or replacing MASt3R depth, or ablating the depth loss) and discuss what 'robust' means when the prior is inaccurate.
  4. [App. B.2, Table 13] The text states that LongSplat 'consistently outperforms COLMAP' on Hike, but Table 13 shows COLMAP with known poses beats LongSplat on forest2 (28.35 vs. 27.87) and university2 (27.25 vs. 26.82). Since COLMAP is the motivating failure case, this comparison should be reported honestly and the claim adjusted. The average is better for LongSplat, but 'consistently' is not supported.
minor comments (4)
  1. [Abstract / Table 11] The wording 'HT-3DGS runs only on T&T (33.53 dB), but falls to 13.75 dB on Free' is internally inconsistent because Table 11 reports a Free result for HT-3DGS. Reword to say HT-3DGS produces very low quality on Free and OOMs on Hike.
  2. [Algorithm 1] The pseudo-code line 'fallback to t' is vague. The text describes a fallback that re-optimizes all past frames globally, but the algorithm should specify which iterations are re-run and how PnP is retried.
  3. [Tables 13/16] Several captions and table headers contain typos ('Qualtitative', 'na¨ıve') and inconsistent use of 'PSNR↑' symbols. Please copyedit the supplementary material.
  4. [§4.3, Table 6] The visibility-adapted window is a central design choice, but the IoU threshold τ is not ablated and the selection rule ('iteratively removing the earliest frame') is described only qualitatively. A short sensitivity experiment or a precise pseudocode statement would help reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; LongSplat uses MASt3R as a learned geometric prior but re-estimates poses, depth, and anchors through photometric and reprojection optimization, so no prediction reduces to its input by construction.

full rationale

The derivation chain is not circular. LongSplat initializes anchors and pose hypotheses from MASt3R (Sec. 3.2, Eq. 8), but the final camera poses are obtained by PnP+RANSAC followed by photometric refinement against the current 3DGS render (Eq. 9) and joint global/local optimization (Sec. 3.4); they are not copies of MASt3R poses. The depth-scale factor in Eq. 10 is a per-frame least-squares scalar computed from the current rendered depth and MASt3R's aligned depth; the depth loss in Eq. 13 therefore constrains the rendered depth to the one-dimensional space spanned by the MASt3R depth vector, a genuine soft regularizer rather than an identity. Anchor unprojection (Eq. 11) inserts MASt3R-depth points into the octree, but those anchors are subsequently trained under multi-view photometric and reprojection losses (Eqs. 14-15), and ablations (Tables 5, 7, 12; Fig. 14) show that removing these components materially changes results, so the output is not a relabeling of the prior. The paper's self-citations appear only in related work and implementation credits; no load-bearing claim relies on an unpublished or author-specific uniqueness theorem. The pose-evaluation limitation acknowledged in App. B.3 ('COLMAP poses are noisy on several Hike videos, so we use the 6 stable sequences') and the use of COLMAP as ground truth on Free are legitimate benchmark-validity concerns that should be weighed as correctness risk, but they do not make the pose estimates equal to COLMAP by construction, so they do not constitute circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method relies on MASt3R as a strong learned prior at multiple stages, plus a set of manually chosen hyperparameters (voxel size, octree thresholds, IoU threshold, λ values). No new physical entities are introduced. The COLMAP-as-ground-truth assumption is a significant and possibly invalid domain assumption that directly affects the reported pose accuracy numbers.

free parameters (4)
  • initial voxel size ε₀ = 0.1
    Chosen for octree anchor formation (Sec. 3.2); not ablations of this value are reported.
  • octree split/prune density thresholds τ_split, τ_prune = 10 and 5 initially, increasing with depth
    Control anchor density and memory; the exact schedule is not specified, preventing exact replication.
  • visibility IoU threshold τ = 0.2
    Defines the local optimization window (Eq. 12); chosen by hand, with only coarse ablations on window size.
  • loss weights λ_depth, λ_reprojection = not reported
    Weights in the total loss (Eq. 15) are never given numerical values, hindering reproduction.
assumptions (4)
  • domain assumption MASt3R provides sufficiently accurate per-frame dense point clouds, correspondences, and aligned depths for the target scenes.
    Initialization (Sec. 3.2), correspondence back-projection (Eq. 8), depth scale alignment (Eq. 10), anchor unprojection (Eq. 11), and depth loss (Eq. 13) all depend on MASt3R outputs.
  • domain assumption Scenes are static with fixed camera intrinsics.
    Stated explicitly in the Limitations section; dynamic objects or zoom changes break the method.
  • domain assumption Anchor-based 3DGS with MLP-decoded Gaussian attributes can fit the target scenes.
    The whole pipeline is built on Scaffold-GS's representation (Sec. 3.1), an assumed capacity from prior work.
  • domain assumption COLMAP poses are a valid ground truth for evaluating pose accuracy.
    Tables 2, 10, 15-16 measure ATE/RPE against COLMAP, yet the paper's motivation is that COLMAP fails on casual videos (Sec. 1, Fig. 2); the authors themselves restrict Hike pose evaluation to 6 stable sequences.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos." pith.science (2026). https://pith.science/paper/7JEEMJZ3

@misc{pith2026250814041,
  author       = {Pith},
  title        = {Pith review of: LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7JEEMJZ3}},
  note         = {Machine review of arXiv:2508.14041}
}
read the original abstract

LongSplat addresses critical challenges in novel view synthesis (NVS) from casually captured long videos characterized by irregular camera motion, unknown camera poses, and expansive scenes. Current methods often suffer from pose drift, inaccurate geometry initialization, and severe memory limitations. To address these issues, we introduce LongSplat, a robust unposed 3D Gaussian Splatting framework featuring: (1) Incremental Joint Optimization that concurrently optimizes camera poses and 3D Gaussians to avoid local minima and ensure global consistency; (2) a robust Pose Estimation Module leveraging learned 3D priors; and (3) an efficient Octree Anchor Formation mechanism that converts dense point clouds into anchors based on spatial density. Extensive experiments on challenging benchmarks demonstrate that LongSplat achieves state-of-the-art results, substantially improving rendering quality, pose accuracy, and computational efficiency compared to prior approaches. Project page: https://linjohnss.github.io/longsplat/

Figures

Figures reproduced from arXiv: 2508.14041 by the authors.

Figure 1
Figure 1. LongSplat achieves robust novel view synthesis from casually captured long videos without provided camera poses. Our approach jointly optimizes camera poses and 3D Gaussian Splatting, producing accurate and visually coherent reconstructions even under challenging conditions. Abstract LongSplat addresses critical challenges in novel view synthe￾sis (NVS) from casually captured long videos characterized by irregular c… view at source ↗
Figure 2
Figure 2. Novel view synthesis for casual long videos. Exist￾ing methods encounter significant challenges when reconstructing scenes from casually captured long videos: COLMAP [51] fails due to incorrect camera pose estimation, CF-3DGS [14] suffers from out-of-memory issues, LocalRF [39] struggles with complex tra￾jectories, and MASt3R [27]+Scaffold-GS [36] provides inaccurate poses leading to degraded rendering quality. In c… view at source ↗
Figure 3
Figure 3. Overview of the LongSplat framework. Given a casually captured long video without known poses, LongSplat incrementally reconstructs the scene through tightly coupled pose estimation and 3D Gaussian Splatting. (a) Initialization converts MASt3R [27] global aligned point cloud into an octree-anchored 3DGS. (b) Global Optimization jointly refines all camera poses and 3D Gaussians for global consistency. (c) Pose estima… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Visualization of our proposed Octree Anchor For￾mation strategy. Given an initial sparse voxelized point cloud, we iteratively perform density-guided adaptive voxel splitting and pruning. Voxels with point cloud density (ρ) exceeding a threshold are split, while those …
Figure 6
Figure 6. Figure 6: Illustration of our Visibility-Adapted Local Window strategy for local optimization. To ensure balanced training of the 3D Gaussians, we dynamically define the optimization window based on anchor visibility overlap. Specifically, we compute the Intersection-over-Union …
Figure 7
Figure 7. Figure 7: Qualitative comparison on the Free dataset [61]. We compare our method with state-of-the-art approaches including NoPe￾NeRF [5], LocalRF [39], CF-3DGS [14], and MASt3R [27] combined with Scaffold-GS [36]. CF-3DGS fails due to memory constraints (OOM), and other baselin…
Figure 8
Figure 8. Figure 8: Visualization of camera trajectories on Free dataset [61]. CF-3DGS [14] encounters OOM and fails for long sequences, whereas our method reliably estimates accurate, stable trajectories, demonstrating superior robustness [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparison on the Tanks and Temples dataset [25]. NoPe-NeRF [5] produces visibly blurred results with inaccurate geometries, while CF-3DGS [14], despite better sharp￾ness, fails to reconstruct fine details accurately. In contrast, our LongSplat method achie…
Figure 10
Figure 10. Figure 10: Qualitative results on the Hike dataset [39]. Compared to existing methods such as LocalRF [39] and MASt3R [27]+Scaffold￾GS [36], our approach significantly improves visual clarity and reconstruction fidelity, accurately capturing complex details and textures in chall…
Figure 11
Figure 11. Figure 11: Robustness analysis on camera pose estimation (Free dataset [61]). We plot cumulative error distributions for ATE, RPE translation, and rotation. Our method consistently achieves lower errors compared to existing methods, demonstrating superior ro￾bustness and reduced…
Figure 12
Figure 12. Figure 12: Qualitative comparison on the CO3Dv2 dataset [46] HT-3DGS Ours HT-3DGS Ours [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Qualitative comparison with HT-3DGS [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Visual comparisons on ablation MASt3R relative pose [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Visual comparisons on ablation studies. The top row shows the camera trajectory estimation and novel view synthesis results when different training components are removed, demonstrating the importance of each proposed module. Removing global optimization, local optimi…
Figure 16
Figure 16. Figure 16: Visualization of camera trajectories on Free dataset [61]. CF-3DGS [14] encounters OOM and fails for long sequences, whereas our method reliably estimates accurate, stable trajectories, demonstrating superior robustness. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: More Qualitative comparison on the Tanks and Temples dataset [25]. NoPe-NeRF [5] produces visibly blurred results with inaccurate geometries, while CF-3DGS [14], despite better sharpness, fails to reconstruct fine details accurately. In contrast, our LongSplat method …
Figure 18
Figure 18. Figure 18: More Qualitative comparison on the Free dataset [61]. We compare our method with state-of-the-art approaches including NoPe-NeRF [5], LocalRF [39], CF-3DGS [14], and MASt3R [27] combined with Scaffold-GS [36]. CF-3DGS fails due to memory constraints (OOM), and other b…
Figure 19
Figure 19. Figure 19: Qualitative results on the Hike dataset [39]. Compared to existing methods such as LocalRF [39] and MASt3R [27]+Scaffold￾GS [36], our approach significantly improves visual clarity and reconstruction fidelity, accurately capturing complex details and textures in chall…
Figure 20
Figure 20. Figure 20: More Qualitative results on the Hike dataset [39]. Compared to existing methods such as LocalRF [39] and MASt3R [27]+Scaffold-GS [36], our approach significantly improves visual clarity and reconstruction fidelity, accurately capturing complex details and textures in …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 66 canonical work pages

  1. [1]

    Hyperreel: High-fidelity 6-dof video with ray- conditioned sampling

    Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollh¨ofer, Johannes Kopf, Matthew O’Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray- conditioned sampling. In CVPR, 2023. 2

  2. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 2

  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 CVPR, 2022

  4. [4]

    Zip-nerf: Anti-aliased grid-based neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2

  5. [5]

    Nope-nerf: Optimising neural radiance field with no pose prior

    Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Vic- tor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In CVPR, 2023. 2, 6, 7, 8, 15, 18

  6. [6]

    Unstructured lumigraph render- ing

    Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph render- ing. In SIGGRAPH, 2001. 2

  7. [7]

    Improving robustness for joint optimization of camera pose and decom- posed low-rank tensorial radiance fields

    Bo-Yu Chen, Wei-Chen Chiu, and Yu-Lun Liu. Improving robustness for joint optimization of camera pose and decom- posed low-rank tensorial radiance fields. In AAAI, 2024. 2

  8. [8]

    View interpolation for image synthesis

    Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In SIGGRAPH, 1993. 2

Show all 77 references
  1. [9]

    Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs

    Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. arXiv preprint arXiv:2306.05410, 2023. 2

  2. [10]

    Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation

    Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv e- prints, 2022. 2

  3. [11]

    Videolifter: Lifting videos to 3d with fast hierarchical stereo alignment, 2025

    Wenyan Cong, Kevin Wang, Jiahui Lei, Colton Stearns, Yuan- hao Cai, Dilin Wang, Rakesh Ranjan, Matt Feiszli, Leonidas Guibas, Zhangyang Wang, Weiyao Wang, and Zhiwen Fan. Videolifter: Lifting videos to 3d with fast hierarchical stereo alignment, 2025. 2

  4. [12]

    Mod- eling and rendering architecture from photographs: A hybrid geometry-and image-based approach

    Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Mod- eling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In SIGGRAPH, 1996. 2

  5. [13]

    Spec- tromotion: Dynamic 3d reconstruction of specular scenes

    Cheng-De Fan, Chen-Wei Chang, Yi-Ruei Liu, Jie-Ying Lee, Jiun-Long Huang, Yu-Chee Tseng, and Yu-Lun Liu. Spec- tromotion: Dynamic 3d reconstruction of specular scenes. In CVPR, 2025. 2

  6. [14]

    Colmap-free 3d gaussian splatting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In CVPR, 2024. 2, 6, 7, 8, 14, 15, 17, 18

  7. [15]

    Fastnerf: High-fidelity neural rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, 2021. 2

  8. [16]

    Multiple view geom- etry in computer vision

    Richard Hartley and Andrew Zisserman. Multiple view geom- etry in computer vision. 2003. 2

  9. [17]

    Automatic photo pop-up

    Derek Hoiem, Alexei A Efros, and Martial Hebert. Automatic photo pop-up. In ACM SIGGRAPH 2005 Papers, 2005. 2

  10. [18]

    Tour into the picture: using a spidery mesh interface to make animation from a single image

    Youichi Horry, Ken-Ichi Anjyo, and Kiyoshi Arai. Tour into the picture: using a spidery mesh interface to make animation from a single image. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques,

  11. [19]

    3d gaussian splatting with grouped uncertainty for unconstrained images

    Hao-Yu Hou, Chia-Chi Hsu, Yu-Chen Huang, Mu-Yi Shen, Wei-Fang Sun, Cheng Sun, Chia-Che Chang, Yu-Lun Liu, and Chun-Yi Lee. 3d gaussian splatting with grouped uncertainty for unconstrained images. In ICASSP, 2025. 2

  12. [20]

    Worldsheet: Wrapping the world in a 3d sheet for view syn- thesis from a single image

    Ronghang Hu, Nikhila Ravi, Alex Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view syn- thesis from a single image. In ICCV, 2020. 2

  13. [21]

    Sfm-free 3d gaussian splatting via hierarchical training

    Bo Ji and Angela Yao. Sfm-free 3d gaussian splatting via hierarchical training. arXiv preprint arXiv:2412.01553, 2024. 2

  14. [22]

    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 TOG, 2023. 2, 3, 6, 7, 15

  15. [23]

    A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets

    Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets. ACM TOG, 2024. 3

  16. [24]

    Infonerf: Ray entropy minimization for few-shot neural volume rendering

    Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2

  17. [25]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM TOG, 2017. 6, 7, 15, 18

  18. [26]

    Cohen, and Richard Szeliski

    Johannes Kopf, Michael F. Cohen, and Richard Szeliski. First- person hyper-lapse videos. ACM TOG, 2014. 2

  19. [27]

    Ground- ing image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing image matching in 3d with mast3r. In ECCV, 2024. 2, 3, 5, 6, 7, 8, 14, 15, 18, 19, 20

  20. [28]

    Virtual view synthesis using backward depth warping algorithm

    Du-Hsiu Li, Hsueh-Ming Hang, and Yu-Lun Liu. Virtual view synthesis using backward depth warping algorithm. In PCS, 2013. 2

  21. [29]

    Mine: Towards continuous depth mpi with nerf for novel view synthesis

    Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. Mine: Towards continuous depth mpi with nerf for novel view synthesis. In ICCV, 2021. 2

  22. [30]

    Genrc: Generative 3d room completion from sparse image collections

    Ming-Feng Li, Yueh-Feng Ku, Hong-Xuan Yen, Chi Liu, Yu- Lun Liu, Albert YC Chen, Cheng-Hao Kuo, and Min Sun. Genrc: Generative 3d room completion from sparse image collections. In ECCV, 2024. 3

  23. [31]

    Barf: Bundle-adjusting neural radiance fields

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2 9

  24. [32]

    Frugalnerf: Fast conver- gence for few-shot novel view synthesis without learned pri- ors

    Chin-Yang Lin, Chung-Ho Wu, Chang-Han Yeh, Shih-Han Yen, Cheng Sun, and Yu-Lun Liu. Frugalnerf: Fast conver- gence for few-shot novel view synthesis without learned pri- ors. CVPR, 2025. 2

  25. [33]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In CVPR, 2024. 3

  26. [34]

    Neural sparse voxel fields

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS,

  27. [35]

    Robust dynamic radiance fields

    Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. InCVPR,

  28. [36]

    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 CVPR, 2024. 2, 3, 4, 6, 7, 8, 14, 15, 18, 19, 20

  29. [37]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713,

  30. [38]

    Humannerf-se: A simple yet effective approach to animate humannerf with diverse poses

    Caoyuan Ma, Yu-Lun Liu, Zhixiang Wang, Wu Liu, Xinchen Liu, and Zheng Wang. Humannerf-se: A simple yet effective approach to animate humannerf with diverse poses. In CVPR,

  31. [39]

    Progressively optimized local radiance fields for robust view synthesis

    Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In CVPR, 2023. 2, 3, 6, 7, 8, 14, 15, 18, 19, 20

  32. [40]

    Nerf: Representing scenes as neural radiance fields for view synthe- sis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthe- sis. Communications of the ACM, 2021. 2

  33. [41]

    Instant neural graphics primitives with a multiresolu- tion hash encoding

    Thomas M¨uller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolu- tion hash encoding. ACM TOG, 2022. 2

  34. [42]

    Orb-slam: a versatile and accurate monocular slam system

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 2015. 2

  35. [43]

    Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs

    Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg- nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2

  36. [44]

    Ct-nerf: Incremental optimizing neural radiance field and poses with complex trajectory.arXiv preprint arXiv:2404.13896, 2024

    Yunlong Ran, Yanxu Li, Qi Ye, Yuchi Huo, Zechun Bai, Jia- hao Sun, and Jiming Chen. Ct-nerf: Incremental optimizing neural radiance field and poses with complex trajectory.arXiv preprint arXiv:2404.13896, 2024. 2

  37. [45]

    Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps

    Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In ICCV, 2021. 2

  38. [46]

    Com- mon objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction

    Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Com- mon objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 12, 13

  39. [47]

    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

  40. [48]

    Free view synthesis

    Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2

  41. [49]

    Stable view synthesis

    Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2

  42. [50]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2

  43. [51]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In CVPR, 2016. 1, 2, 6

  44. [52]

    Driveenv-nerf: Exploration of a nerf-based autonomous driving environment for real-world performance validation

    Mu-Yi Shen, Chia-Chi Hsu, Hao-Yu Hou, Yu-Chen Huang, Wei-Fang Sun, Chia-Che Chang, Yu-Lun Liu, and Chun-Yi Lee. Driveenv-nerf: Exploration of a nerf-based autonomous driving environment for real-world performance validation. arXiv preprint arXiv:2403.15791, 2024. 2

  45. [53]

    Boostmvsnerfs: Boosting mvs-based nerfs to generalizable view synthesis in large-scale scenes

    Chih-Hai Su, Chih-Yao Hu, Shr-Ruei Tsai, Jie-Ying Lee, Chin-Yang Lin, and Yu-Lun Liu. Boostmvsnerfs: Boosting mvs-based nerfs to generalizable view synthesis in large-scale scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 3

  46. [54]

    Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 2

  47. [55]

    Fed3dgs: Scalable 3d gaussian splatting with federated learning

    Teppei Suzuki. Fed3dgs: Scalable 3d gaussian splatting with federated learning. arXiv preprint arXiv:2403.11460, 2024. 3

  48. [56]

    Vi- sual slam algorithms: A survey from 2010 to 2016

    Takafumi Taketomi, Hideaki Uchiyama, and Sei Ikeda. Vi- sual slam algorithms: A survey from 2010 to 2016. IPSJ Transactions on Computer Vision and Applications, 2017. 2

  49. [57]

    Block-nerf: Scalable large scene neural view synthesis

    Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Prad- han, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In CVPR, 2022. 3

  50. [58]

    Single-view view synthe- sis with multiplane images

    Richard Tucker and Noah Snavely. Single-view view synthe- sis with multiplane images. In CVPR, 2020. 2

  51. [59]

    Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields

    Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2

  52. [60]

    3d reconstruction with spatial memory

    Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 3

  53. [61]

    F2- nerf: Fast neural radiance field training with free camera trajectories

    Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2- nerf: Fast neural radiance field training with free camera trajectories. In CVPR, 2023. 6, 7, 8, 15, 17, 18

  54. [62]

    Continuous 3d perception model with persistent state

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387,

  55. [63]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 3

  56. [64]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 6 10

  57. [65]

    NeRF −−: Neural radiance fields without known camera parameters

    Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF −−: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2

  58. [66]

    Sin- erf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction

    Yitong Xia, Hao Tang, Radu Timofte, and Luc Van Gool. Sin- erf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. 2022. 2

  59. [67]

    Sinnerf: Training neural radiance fields on complex scenes from a single image

    Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. InECCV, 2022. 2

  60. [68]

    Point-nerf: Point- based neural radiance fields

    Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point- based neural radiance fields. In CVPR, 2022. 2

  61. [69]

    Freenerf: Im- proving few-shot neural rendering with free frequency regu- larization

    Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Im- proving few-shot neural rendering with free frequency regu- larization. In CVPR, 2023. 2

  62. [70]

    Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass

    Jianing Yang, Alexander Sax, Kevin J Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. arXiv preprint arXiv:2501.13928, 2025. 3

  63. [71]

    inerf: Inverting neural radiance fields for pose estimation

    Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. 2

  64. [72]

    Plenoctrees for real-time rendering of neural radiance fields

    Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2

  65. [73]

    Cat-3dgs: A context-adaptive triplane approach to rate- distortion-optimized 3dgs compression

    Yu-Ting Zhan, Cheng-Yuan Ho, Hebi Yang, Yi-Hsin Chen, Jui Chiu Chiang, Yu-Lun Liu, and Wen-Hsiao Peng. Cat-3dgs: A context-adaptive triplane approach to rate- distortion-optimized 3dgs compression. arXiv preprint arXiv:2503.00357, 2025. 2

  66. [74]

    Nerf++: Analyzing and improving neural radiance fields

    Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv:2010.07492, 2020. 2

  67. [75]

    Differentiable point-based radiance fields for efficient view synthesis

    Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. InSIGGRAPH Asia 2022 Conference Papers, 2022. 2

  68. [76]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6

  69. [77]

    Stereo magnification: Learning view syn- thesis using multiplane images

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view syn- thesis using multiplane images. 2018. 2 11 A. Implementation Details We implement LongSplat using PyTorch. Our rendering and 3D Gaussian updates are accelerated u...

Pith tools

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