Pith. sign in

REVIEW 3 major objections 3 minor 59 references

A unified Gaussian surfel representation lets an active scanner detect back-facing, under-reconstructed surfaces and plan a short path that yields complete, photorealistic object models in minutes.

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 11:11 UTC pith:DYH2K6WP

load-bearing objection Solid, well-engineered active reconstruction paper with a genuinely new NBP + geometry-aware uncertainty combination; the empirical case is strong, but Eq. 13's unbounded confidence update can break the planner as written and needs a stated clamp/normalization. the 3 major comments →

arxiv 2601.06997 v2 pith:DYH2K6WP submitted 2026-01-11 cs.RO cs.CV

ObjSplat: Geometry-Aware Gaussian Surfels for Active Object Reconstruction

Yuetao Li , Zhizhou Jia , Yu Zhang , Qun Hao , Shaohui Zhang This is my paper
classification cs.RO cs.CV
keywords active object reconstructionGaussian surfelsnext-best-path planningback-face visibilityocclusion-aware covisibilityview planningRGB-D perception3D 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.

ObjSplat tries to show that autonomous high-fidelity object reconstruction can be made both faster and more complete by coupling the reconstruction representation itself with the view planner. The paper argues that opacity- or depth-based uncertainty cues are unreliable for object-centric scanning, and replaces them with a geometry-aware evaluation that flags back-facing surfaces and uses occlusion-aware covisibility. It then replaces greedy next-best-view selection with a next-best-path planner that performs multi-step lookahead on a graph of candidate viewpoints, trading information gain against travel distance. The claimed result is physically consistent, watertight meshes and photorealistic Gaussian models within minutes, with path lengths about a quarter of those of greedy planners.

Core claim

The central claim is that 2D Gaussian surfels — flat, disk-like primitives with a zero-thickness scale — can serve as a single representation for both incremental reconstruction and exploration guidance in active object scanning. The paper introduces a viewpoint evaluation pipeline that renders normal, depth, and opacity maps from the current model and marks pixels where the rendered normal faces the camera (back-face), where opacity is low, or where depth is geometrically inconsistent, and it formulates an occlusion-aware covisibility ratio that corrects for self-occlusion and back-facing geometry. Uncertainty maps built from these signals drive a next-best-path planner that solves a prize-

What carries the argument

The load-bearing object is the Gaussian surfel — a 2D Gaussian disk with zero z-scale, whose rotation quaternion directly yields a surface normal. Two mechanisms built on it carry the argument: (1) the back-face mask m_back(u) = [Nhat(u)·z > 0], which classifies any rendered patch whose normal points toward the camera as an under-reconstructed defect, and (2) the next-best-path planner, which constructs a k-nearest-neighbor graph over candidate viewpoints, assigns edge weights w = d_ij / (α + β(U_i + U_j)), and selects among the top-M shortest paths the one maximizing normalized information gain minus travel cost. The covisibility check uses the same rendered depth and normal maps to reject

Load-bearing premise

The planner treats any rendered surface patch whose normal points toward the camera as an under-reconstructed defect, so if the partial model contains misoriented or inward-facing normals, the robot will waste motion chasing phantom back-face regions.

What would settle it

Run the next-best-path pipeline on a hollow thin-walled object whose sparse-view optimization is known to produce flipped normals, and compare the planned paths and final completion against a version where the rendered normals are replaced with ground-truth orientations; a substantial divergence in path length or completion ratio would show the back-face signal is driving phantom exploration.

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

If this is right

  • An autonomous robot can reconstruct an unknown object to over 90% surface completion in about 30 views, with the next-best-path variant doing so in under 4 meters of travel.
  • Greedy next-best-view planners pay a large hidden cost: comparable quality is achievable, but at roughly four to five times the path length, so movement cost belongs inside the view-selection objective.
  • Opacity- or depth-residual-based uncertainty cues mislead planning on object-centric, non-closed surfaces; back-face visibility and occlusion-aware covisibility are necessary to avoid premature termination on self-occluded and thin structures.
  • Because the Gaussian surfel representation is shared across mapping, uncertainty evaluation, and offline refinement, the whole loop runs with millisecond-scale uncertainty rendering and an end-to-end mission time of roughly five minutes.
  • Executing a full planned path outperforms a receding-horizon replanning strategy, because frequent replanning disrupts trajectory smoothness and adds overhead without improving the final model.

Where Pith is reading between the lines

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

  • If the back-face signal proves reliable across object classes, the same planning machinery could be lifted onto other surface-based representations (TSDF, meshes) by simply rendering normals and depths, making the approach representation-agnostic rather than tied to Gaussian surfels.
  • The dramatic path-length reduction suggests that information-gain-only objectives systematically overvalue distant views; adding traversability or reachability costs to the edge weights could extend the gains to cluttered or constrained workspaces.
  • Because the method assumes the object can be cleanly segmented in every frame, its behavior on non-segmentable or multi-object scenes is an open question; a testable extension would be to fall back on saliency or foreground-proposal masks and measure the planning degradation.
  • The dense uncertainty maps render in roughly 2 ms, which is fast enough to re-plan at a higher rate than once per full path; interleaving partial re-planning with arm motion could recover some of the adaptivity of greedy planners without their travel cost.

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 / 3 minor

Summary. ObjSplat is an active object reconstruction system that uses 2D Gaussian surfels as a unified representation for incremental reconstruction and for guiding next-view selection. The framework maintains a GSurfel model from RGB-D frames, evaluates candidate viewpoints with a geometry-aware uncertainty map that combines surfel confidence, back-face detection, and visibility-based completeness, and plans multi-step trajectories with a next-best-path (NBP) planner over a dynamically built k-NN graph. The paper reports simulations on 16 GSO objects against five baselines, ablations of the main components, robustness to initial pose, and real-world experiments on four cultural artifacts. The headline result is that the NBP variant reaches 91.42% completion with 3.96 m of movement after 30 views, versus 91.89% at 18.02 m for the greedy NBV variant, while achieving best or near-best PSNR, Chamfer Distance, and F-Score.

Significance. If the results hold, the paper makes a practically important contribution: a non-greedy trajectory planner that reduces movement cost by roughly 4–5x relative to greedy NBV while maintaining comparable reconstruction quality, and a geometry-aware uncertainty metric that explicitly handles back-face and occlusion effects in object-centric scanning. The evaluation is broad by the standards of the field: five baselines, 16-object simulation suite, component ablations, initial-pose robustness trials, and four real-world artifacts. The paper does not claim machine-checked proofs and does not provide code, so the contribution rests on the reported experiments. The main risk is not circularity but internal consistency: the confidence update as written can invalidate the shortest-path planner, so the headline efficiency result is not yet fully supported by the equations as stated.

major comments (3)
  1. [III-B3, Eqs. (13), (16), (17)] The confidence model is internally inconsistent. Eq. (13) states κ_i ∈ [0,1] but updates κ_i ← γ_i exp(β_i), where γ_i is a sum over all valid historical views of w_ij·max(0,n_i^T v_ij). Each summand is ≤1, so γ_i grows approximately linearly with the number of observations (roughly 10–30 by view 30), and exp(β_i)≥1. No clamping or normalization is specified. The rendered confidence map K(u) in Eq. (16) is a blend of these κ_i values, so K(u) can exceed 1, making U(u)=λ_k(1−K(u))+λ_bB(u)+λ_vV(u) possibly negative. Consequently, the edge weights in Eq. (17), w(v_i,v_j)=d_ij/[α+β(U(v_i)+U(v_j))], can become negative whenever the denominator is negative. Yen's k-shortest-simple-paths algorithm (ref [57]) requires non-negative edge weights. As written, the NBP planner's objective is not well-posed. Please specify the normalization/clipping of κ_i, K, and U, or state explicitly that the imple
  2. [Tables I and II, Sec. IV-C2] The headline quantitative comparisons are reported as single aggregate numbers without variance or per-object statistics. For example, at convergence Ours-NBP has CR 91.42% versus Ours-NBV at 91.89% (a difference of 0.47 pp), and CD 0.611 mm versus 0.609 mm; without standard deviations or paired per-object comparisons it is impossible to assess whether these differences are meaningful or whether NBP truly 'matches' NBV. Fig. 8 shows shaded error bands but the numerical tables do not. Please report mean±std over the 16 objects for the metrics in Tables I and II, and ideally paired significance tests for the key CR/CD/PSNR comparisons.
  3. [Sec. III-C2, III-C3, Eq. (16)] The dynamic weighting schedule and stopping criterion are under-specified. Section III-C3 says that once visibility-based uncertainty stabilizes, λ_v is set to 0. Section III-C2 says termination uses Û_q(u)=λ_bB(u)+λ_vV(u). If λ_v=0 in the refinement phase, the stopping criterion no longer depends on visibility or confidence, only on back-face area B; low-confidence front-facing regions would not influence termination. The threshold τ_stop and the rule for when 'visibility stabilizes' are also not defined. This is not merely cosmetic: it affects the claimed completeness results. Please give the exact schedule and stopping rule.
minor comments (3)
  1. [Eq. (13)] The notation is confusing: the text says β_i = 1−∥µ_i∥ is computed from the 'mean viewing direction' of S_valid(g_i), but µ_i was previously defined as the surfel center. If µ_i is a center position, its norm is a distance from the origin, not an angular diversity; if a mean direction is intended, a different symbol and normalization should be used.
  2. [Table I] The table formatting in the provided manuscript is broken: the Test-row values for FisherRF and GauSS-MI appear to be misaligned, with missing or merged entries (e.g., 'Test 25.43 0.9210.101 7.821 28.02 ...'). Please reformat so each column is legible.
  3. [Sec. IV-B] Typo: 'Evaluation Metircs' should be 'Evaluation Metrics'.

Circularity Check

0 steps flagged

No significant circularity; central claims are evaluated against external ground truth and independent baselines.

full rationale

ObjSplat's reconstruction and planning pipeline is an engineering system, not a derivation that reduces to its inputs. The progressive update mask (Eq. 7) and uncertainty map (Eq. 16) are defined in terms of the current Gaussian surfel model, but this is standard closed-loop active perception: the rendered cues guide view selection, and the resulting reconstruction quality is measured externally against GSO ground-truth meshes and real-world artifacts. The NBP planner's reported efficiency gains are supported by quantitative comparisons with independent baselines (SEE, PB-NBV, MA-SCVP, FisherRF, GauSS-MI) and by ablations; there is no fitted parameter that is renamed as a prediction. The self-citations [21] (PB-NBV) and [28] (ActiveSplat) are prior work used as a comparison baseline and a heuristic to critique, respectively; they are not load-bearing assumptions and no uniqueness claim is imported from them. The paper's own limitation statement (single static rigid objects; challenging optical properties) is a scope limitation, not evidence of circularity. The skeptical concern that the confidence update in Eq. 13 is unclamped and can make Eq. 17 edge weights negative is a correctness/under-specification risk, not a circularity, and is outside this pass.

Axiom & Free-Parameter Ledger

9 free parameters · 7 axioms · 0 invented entities

The central claim is an empirical system claim. No free parameters are explicitly fitted to the headline metrics, but several hand-set thresholds directly shape the uncertainty signal that drives planning and stopping. The most consequential assumption is the back-face heuristic (Eq. 7d), which treats rendered normals facing the camera as proof of incompleteness; on thin, hollow, or self-occluded geometry this is only as trustworthy as the current surfel model.

free parameters (9)
  • tau_O (low-opacity threshold) = 0.5
    Eq. 7a: pixels with rendered opacity below 0.5 are treated as empty/incomplete. Hand-set, no sensitivity analysis.
  • tau_C (photometric discrepancy threshold) = 0.25
    Eq. 7b: MSE threshold for adding surfels to texture-mismatched pixels. Hand-set.
  • lambda (geometric deviation factor) = 2
    Eq. 7c: rendered depth must exceed observed depth by 2× MDE to be a defect. Hand-set.
  • tau_d (depth tolerance in covisibility) = unspecified
    Eq. 10: invalidates reprojected correspondences whose occluded depth difference exceeds tau_d. Value not reported.
  • alpha, beta (edge-weight constants) = unspecified
    Eq. 17: balance distance against uncertainty reward in the graph; values not reported.
  • path-score lambda (exploration vs efficiency) = unspecified
    Eq. 18: weights collected uncertainty against path length in path selection; no value or schedule is given.
  • c_0, tau (front-view sigmoid parameters) = 0.5, 0.1
    Eq. 12b: prioritizes near-frontal observations at about 60 degrees with temperature 0.1. Hand-set.
  • d_max (distance penalty scale) = unspecified
    Eq. 12a: max distance for geometric weight; not specified.
  • lambda_k, lambda_b, lambda_v (dynamic uncertainty weights) = 1/1/1 early, lambda_v=0 later
    Eq. 16 and Sec. III-C3: the schedule is described qualitatively ('once visibility-based uncertainty stabilizes'), not as a precise rule.
axioms (7)
  • standard math Differentiable Gaussian splatting and alpha-blending correctly model rendering of 2D Gaussian surfels.
    Relied on throughout Sec. III-A for rendered color, depth, normal, opacity, and gradient-based optimization.
  • domain assumption Surfel normals (from quaternion and depth finite differences) are accurate enough to drive update masks and uncertainty.
    Back-face detection (Eq. 7d), normal losses (Eq. 4), and the uncertainty map (Eq. 16) all trust normals on partially reconstructed models.
  • domain assumption SAM2 masking and RGB-D depth are correct within the object mask.
    All masks, back-projections (Eq. 8), and covisibility sampling assume clean foreground segmentation and valid sensor depth; no failure analysis is provided.
  • ad hoc to paper The back-face heuristic m_back(u):=[Nhat(u)·z>0] reliably identifies under-reconstructed regions.
    Eq. 7d treats rendered normals pointing at the camera as evidence of incompleteness. This is a modeling assumption unique to this paper, not a derived fact.
  • ad hoc to paper The weighted uncertainty map (Eq. 16) and its stopping criterion reflect true reconstruction completeness.
    The planner and termination rule trust this decomposition; no calibration against ground truth is presented beyond qualitative visuals.
  • domain assumption Robot-arm/turntable calibration and object-centric pose tracking remain valid during scanning.
    Sec. III-D1 relies on pairwise point-cloud registration to correct pose drift; registration failure or non-rigid motion would break model fusion and planning.
  • standard math Yen's top-M shortest simple paths on the k-NN graph provide a sufficient search space for globally efficient trajectories.
    Sec. III-C2 uses graph connectivity and Yen's algorithm without proof that the k-NN graph preserves near-optimal paths.

pith-pipeline@v1.3.0-alltime-deepseek · 24505 in / 12271 out tokens · 129096 ms · 2026-08-03T11:11:51.667560+00:00 · methodology

0 comments
read the original abstract

Autonomous high-fidelity object reconstruction is fundamental for creating digital assets and bridging the simulation-to-reality gap in robotics. We present ObjSplat, an active reconstruction framework that leverages Gaussian surfels as a unified representation to progressively reconstruct unknown objects with both photorealistic appearance and accurate geometry. Addressing the limitations of conventional opacity or depth-based cues, we introduce a geometry-aware viewpoint evaluation pipeline that explicitly models back-face visibility and occlusion-aware multi-view covisibility, reliably identifying under-reconstructed regions even on geometrically complex objects. Furthermore, to overcome the limitations of greedy planning strategies, ObjSplat employs a next-best-path (NBP) planner that performs multi-step lookahead on a dynamically constructed spatial graph. By jointly optimizing information gain and movement cost, this planner generates globally efficient trajectories. Extensive experiments in simulation and on real-world cultural artifacts demonstrate that ObjSplat produces physically consistent models within minutes, achieving superior reconstruction fidelity and surface completeness while significantly reducing scan time and path length compared to state-of-the-art approaches. Project page: https://li-yuetao.github.io/ObjSplat-page/ .

Figures

Figures reproduced from arXiv: 2601.06997 by Qun Hao, Shaohui Zhang, Yuetao Li, Yu Zhang, Zhizhou Jia.

Figure 1
Figure 1. Figure 1: ObjSplat autonomously plans viewpoints and progressively recon [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System Overview. ObjSplat progressively reconstructs unknown objects from RGB-D frames using Gaussian surfels as a unified representation. (Top left) Incoming frames are fused into the global model, where geometry–texture joint optimization, enforcing both photometric and geometric consistency. (Right) A geometry-aware view evaluation pipeline renders an uncertainty map by integrating occlusion-aware covis… view at source ↗
Figure 4
Figure 4. Figure 4: Co-visible views comparison. While conventional frustum-based metrics (Left) fail to account for self-occlusion, leading to false correspon￾dences, our rendering-based check (Right) explicitly detects occlusion and back-facing surfaces, thus correctly identifying valid and invalid co-visible points, ensuring valid geometric constraints and confidence updating. surfels maintain a consistent spatial distribu… view at source ↗
Figure 5
Figure 5. Figure 5: Uncertainty qualification comparison on the [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Experimental Platforms. (Left) The Gazebo simulation environment provides realistic rendering and physics-based robot models for quantitative evaluation. (Right) The real-world platform comprises an Elfin-05 robotic arm equipped with a calibrated industrial 3D camera. The system architecture is built on modular ROS nodes for Gaussian mapping, NBP planning, and visualization across both domains. the SH degr… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of reconstruction results on six representative objects from the GSO dataset. ObjSplat recovers sharp textures and clean [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Quantitative comparison of reconstruction progress over the number of views (top row) and path length (bottom row). We report novel view PSNR, [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual comparison of reconstruction completeness and exploration efficiency on the [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Reconstruction quality comparison on the [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Real-World Progressive Reconstruction Process Results. We show four snapshots of the NBP planning execution. In each step, the robot executes a path to a sub-goal with high estimated uncertainty (visualized in the inset). The final column shows the complete trajectory, total path length, and the final PSNR (dB) of the training views. TABLE VI IMPACT OF THE OBJECT’S INITIAL POSE. (30 VIEWS AND 5 TRIALS) Ob… view at source ↗
Figure 12
Figure 12. Figure 12: Final reconstruction results of real-world artifacts. We display the rendered images from the final Gaussian surfels (left) and the corresponding extracted surface meshes (right). Our method recovers high-fidelity texture and geometry even for objects with complex topology and surface patterns. uncertainty regions (red/bright areas) and generates collision￾free paths for exploration. As reconstruction pro… 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

59 extracted references · 5 linked inside Pith

  1. [1]

    Metascenes: Towards automated replica creation for real- world 3d scans,

    H. Yu, B. Jia, Y . Chen, Y . Yang, P. Li, R. Su, J. Li, Q. Li, W. Liang, S.-C. Zhuet al., “Metascenes: Towards automated replica creation for real- world 3d scans,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 1667–1679

    2025

  2. [2]

    Digital twin catalog: A large-scale photorealistic 3d object digital twin dataset,

    Z. Dong, K. Chen, Z. Lv, H.-X. Yu, Y . Zhang, C. Zhang, Y . Zhu, S. Tian, Z. Li, G. Moffattet al., “Digital twin catalog: A large-scale photorealistic 3d object digital twin dataset,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 753–763

    2025

  3. [3]

    What is the best 3d scene representation for robotics? from geometric to foundation models,

    T. Deng, Y . Pan, S. Yuan, D. Li, C. Wang, M. Li, L. Chen, L. Xie, D. Wang, J. Wanget al., “What is the best 3d scene representation for robotics? from geometric to foundation models,”arXiv preprint arXiv:2512.03422, 2025

    arXiv 2025

  4. [4]

    Reconciling reality through simulation: A real- to-sim-to-real approach for robust manipulation,

    M. Torne, A. Simeonov, Z. Li, A. Chan, T. Chen, A. Gupta, and P. Agrawal, “Reconciling reality through simulation: A real- to-sim-to-real approach for robust manipulation,”arXiv preprint arXiv:2403.03949, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Scalable real2sim: Physics-aware asset generation via robotic pick-and-place setups,

    N. Pfaff, E. Fu, J. Binagia, P. Isola, and R. Tedrake, “Scalable real2sim: Physics-aware asset generation via robotic pick-and-place setups,”arXiv preprint arXiv:2503.00370, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Real2render2real: Scaling robot data without dynamics simulation or robot hardware,

    J. Yu, L. Fu, H. Huang, K. El-Refai, R. A. Ambrus, R. Cheng, M. Z. Irshad, and K. Goldberg, “Real2render2real: Scaling robot data without dynamics simulation or robot hardware,”arXiv preprint arXiv:2505.09601, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Gaussgym: An open-source real-to-sim framework for learning locomotion from pixels,

    A. Escontrela, J. Kerr, A. Allshire, J. Frey, R. Duan, C. Sferrazza, and P. Abbeel, “Gaussgym: An open-source real-to-sim framework for learning locomotion from pixels,”arXiv preprint arXiv:2510.15352, 2025

    arXiv 2025

  8. [8]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Commun. ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  9. [9]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Trans. Graphics, vol. 42, no. 4, pp. 1–14, 2023

    2023

  10. [10]

    Structure-from-motion revisited,

    J. L. Sch ¨onberger and J.-M. Frahm, “Structure-from-motion revisited,” in IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  11. [11]

    Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields,

    S. Lee, L. Chen, J. Wang, A. Liniger, S. Kumar, and F. Yu, “Uncertainty guided policy for active robotic 3d reconstruction using neural radiance fields,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 12 070– 12 077, 2022

    2022

  12. [12]

    Activenerf: Learning where to see with uncertainty estimation,

    X. Pan, Z. Lai, S. Song, and G. Huang, “Activenerf: Learning where to see with uncertainty estimation,” inEuropean Conf. on Computer Vision (ECCV). Springer, 2022, pp. 230–246

    2022

  13. [13]

    Active implicit object reconstruction using uncertainty-guided next-best-view optimization,

    D. Yan, J. Liu, F. Quan, H. Chen, and M. Fu, “Active implicit object reconstruction using uncertainty-guided next-best-view optimization,” IEEE Robotics and Automation Letters, vol. 8, no. 10, pp. 6395–6402, 2023

    2023

  14. [14]

    Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering,

    L. Jin, X. Chen, J. R ¨uckin, and M. Popovi ´c, “Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering,” inIEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 11 305–11 312

    2023

  15. [15]

    The surface edge explorer (see): A measurement-direct approach to next best view planning,

    R. Border and J. D. Gammell, “The surface edge explorer (see): A measurement-direct approach to next best view planning,”Intl. J. of Robotics Research, vol. 43, no. 10, pp. 1506–1532, 2024

    2024

  16. [16]

    Pvp-recon: Progressive view planning via warping consistency for sparse-view surface reconstruction,

    S. Ye, Y . He, M. Lin, J. Sheng, R. Fan, Y . Han, Y . Hu, R. Yi, Y .-H. Wen, Y .-J. Liuet al., “Pvp-recon: Progressive view planning via warping consistency for sparse-view surface reconstruction,”ACM Trans. Graphics, vol. 43, no. 6, pp. 1–13, 2024

    2024

  17. [17]

    Fisherrf: Active view selection and mapping with radiance fields using fisher information,

    W. Jiang, B. Lei, and K. Daniilidis, “Fisherrf: Active view selection and mapping with radiance fields using fisher information,” inEuropean Conf. on Computer Vision (ECCV). Springer, 2024, pp. 422–440

    2024

  18. [18]

    Neural visibility field for uncertainty-driven active mapping,

    S. Xue, J. Dill, P. Mathur, F. Dellaert, P. Tsiotra, and D. Xu, “Neural visibility field for uncertainty-driven active mapping,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 18 122–18 132

    2024

  19. [19]

    Gennbv: Generalizable next-best-view policy for active 3d reconstruction,

    X. Chen, Q. Li, T. Wang, T. Xue, and J. Pang, “Gennbv: Generalizable next-best-view policy for active 3d reconstruction,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 436– 16 445

    2024

  20. [20]

    Gauss-mi: Gaussian splatting shannon mutual information for active 3d reconstruction,

    Y . Xie, Y . Cai, Y . Zhang, L. Yang, and J. Pan, “Gauss-mi: Gaussian splatting shannon mutual information for active 3d reconstruction,”arXiv preprint arXiv:2504.21067, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Pb-nbv: Efficient projection- based next-best-view planning framework for reconstruction of unknown objects,

    Z. Jia, Y . Li, Q. Hao, and S. Zhang, “Pb-nbv: Efficient projection- based next-best-view planning framework for reconstruction of unknown objects,”IEEE Robotics and Automation Letters, 2025

    2025

  22. [22]

    Bayesian nerf: Quantifying un- certainty with volume density for neural implicit fields,

    S. Lee, K. Kang, S. Ha, and H. Yu, “Bayesian nerf: Quantifying un- certainty with volume density for neural implicit fields,”IEEE Robotics and Automation Letters, 2025

    2025

  23. [23]

    Scvp: Learning one-shot view planning via set covering for unknown object reconstruction,

    S. Pan, H. Hu, and H. Wei, “Scvp: Learning one-shot view planning via set covering for unknown object reconstruction,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 1463–1470, 2022

    2022

  24. [24]

    Integrating one-shot view planning with a single next- best view via long-tail multiview sampling,

    S. Pan, H. Hu, H. Wei, N. Dengler, T. Zaenker, M. Dawood, and M. Bennewitz, “Integrating one-shot view planning with a single next- best view via long-tail multiview sampling,”IEEE Trans. Robotics, 2024

    2024

  25. [25]

    Exploiting priors from 3d diffusion models for rgb-based one-shot view planning,

    S. Pan, L. Jin, X. Huang, C. Stachniss, M. Popovi ´c, and M. Bennewitz, “Exploiting priors from 3d diffusion models for rgb-based one-shot view planning,” inIEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 13 341–13 348

    2024

  26. [26]

    Dm-osvp++: One-shot view planning using 3d diffusion models for active rgb-based object reconstruction,

    ——, “Dm-osvp++: One-shot view planning using 3d diffusion models for active rgb-based object reconstruction,”arXiv, 2025

    2025

  27. [27]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 18 039–18 048

    2024

  28. [28]

    Activesplat: High-fidelity scene reconstruction through active gaussian splatting,

    Y . Li, Z. Kuang, T. Li, Q. Hao, Z. Yan, G. Zhou, and S. Zhang, “Activesplat: High-fidelity scene reconstruction through active gaussian splatting,”IEEE Robotics and Automation Letters, 2025

    2025

  29. [29]

    Pixelwise view selection for unstructured multi-view stereo,

    J. L. Sch ¨onberger, E. Zheng, M. Pollefeys, and J.-M. Frahm, “Pixelwise view selection for unstructured multi-view stereo,” inEuropean Conf. on Computer Vision (ECCV), 2016

    2016

  30. [30]

    Screened poisson surface reconstruction,

    M. Kazhdan and H. Hoppe, “Screened poisson surface reconstruction,” ACM Trans. Graphics, vol. 32, no. 3, pp. 1–13, 2013

    2013

  31. [31]

    Seamless texture optimization for rgb-d reconstruction,

    Y . Fu, Q. Yan, J. Liao, H. Zhou, J. Tang, and C. Xiao, “Seamless texture optimization for rgb-d reconstruction,”IEEE Trans. Visualization and Comp. Graphics, vol. 29, no. 3, pp. 1845–1859, 2021

    2021

  32. [32]

    Adaptive joint optimization for 3d reconstruction with differentiable rendering,

    J. Zhang, Z. Wan, and J. Liao, “Adaptive joint optimization for 3d reconstruction with differentiable rendering,”IEEE Trans. Visualization and Comp. Graphics, vol. 29, no. 6, pp. 3039–3051, 2022

    2022

  33. [33]

    Neus: learning neural implicit surfaces by volume rendering for multi-view reconstruction,

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: learning neural implicit surfaces by volume rendering for multi-view reconstruction,” inAdvances in Neural Information Processing Systems (NIPS), 2021, pp. 27 171–27 183

    2021

  34. [34]

    Neuralangelo: High-fidelity neural surface reconstruction,

    Z. Li, T. M ¨uller, A. Evans, R. H. Taylor, M. Unberath, M.-Y . Liu, and C.- H. Lin, “Neuralangelo: High-fidelity neural surface reconstruction,” in IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 8456–8465

    2023

  35. [35]

    Nerfprior: Learning neural radiance field as a prior for indoor scene reconstruction,

    W. Zhang, E. Y .-t. Jia, J. Zhou, B. Ma, K. Shi, Y .-S. Liu, and Z. Han, “Nerfprior: Learning neural radiance field as a prior for indoor scene reconstruction,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 11 317–11 327

    2025

  36. [36]

    Sparserecon: Neural implicit surface reconstruction from sparse views with feature and depth consistencies,

    L. Han, X. Zhang, H. Song, K. Shi, Y .-S. Liu, and Z. Han, “Sparserecon: Neural implicit surface reconstruction from sparse views with feature and depth consistencies,” inIntl. Conf. on Computer Vision (ICCV), 2025, pp. 28 514–28 524

    2025

  37. [37]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” inSIGGRAPH, 2024, pp. 1– 11

    2024

  38. [38]

    Pgsr: Planar-based gaussian splatting for effi- cient and high-fidelity surface reconstruction,

    D. Chen, H. Li, W. Ye, Y . Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “Pgsr: Planar-based gaussian splatting for effi- cient and high-fidelity surface reconstruction,”IEEE Trans. Visualization and Comp. Graphics, 2024

    2024

  39. [39]

    High-quality surface reconstruction using gaussian surfels,

    P. Dai, J. Xu, W. Xie, X. Liu, H. Wang, and W. Xu, “High-quality surface reconstruction using gaussian surfels,” inSIGGRAPH, 2024, pp. 1–11

    2024

  40. [40]

    Quadratic gaussian splatting: High quality surface reconstruction with second-order geometric primitives,

    Z. Zhang, B. Huang, H. Jiang, L. Zhou, X. Xiang, and S. Shen, “Quadratic gaussian splatting: High quality surface reconstruction with second-order geometric primitives,” inIntl. Conf. on Computer Vision (ICCV), 2025, pp. 28 260–28 270

    2025

  41. [41]

    Gs-sdf: Lidar- augmented gaussian splatting and neural sdf for geometrically consistent rendering and reconstruction,

    J. Liu, Y . Wan, B. Wang, C. Zheng, J. Lin, and F. Zhang, “Gs-sdf: Lidar- augmented gaussian splatting and neural sdf for geometrically consistent rendering and reconstruction,” inIEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 19 391–19 398

    2025

  42. [42]

    Gs- livo: Real-time lidar, inertial, and visual multi-sensor fused odometry with gaussian mapping,

    S. Hong, C. Zheng, Y . Shen, C. Li, F. Zhang, T. Qin, and S. Shen, “Gs- livo: Real-time lidar, inertial, and visual multi-sensor fused odometry with gaussian mapping,”IEEE Trans. Robotics, 2025

    2025

  43. [43]

    Milo: Mesh-in-the-loop gaussian splatting for detailed and efficient surface reconstruction,

    A. Gu ´edon, D. Gomez, N. Maruani, B. Gong, G. Drettakis, and M. Ovsjanikov, “Milo: Mesh-in-the-loop gaussian splatting for detailed and efficient surface reconstruction,”ACM Trans. Graphics, vol. 44, no. 6, pp. 1–15, 2025

    2025

  44. [44]

    Triangle splatting for real-time radiance field rendering,

    J. Held, R. Vandeghen, A. Deliege, A. Hamdi, A. Cioppa, S. Giancola, A. Vedaldi, B. Ghanem, A. Tagliasacchi, and M. Van Droogenbroeck, “Triangle splatting for real-time radiance field rendering,”arXiv, 2025. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 17

    2025

  45. [45]

    Meshsplatting: Differentiable rendering with opaque meshes,

    J. Held, S. Son, R. Vandeghen, D. Rebain, M. Gadelha, Y . Zhou, A. Cioppa, M. C. G Lin, M. Van Droogenbroeck, and A. Tagliasacchi, “Meshsplatting: Differentiable rendering with opaque meshes,”arXiv, 2025

    2025

  46. [46]

    Gaussianobject: High-quality 3d object reconstruction from four views with gaussian splatting,

    C. Yang, S. Li, J. Fang, R. Liang, L. Xie, X. Zhang, W. Shen, and Q. Tian, “Gaussianobject: High-quality 3d object reconstruction from four views with gaussian splatting,”ACM Trans. Graphics, vol. 43, no. 6, pp. 1–13, 2024

    2024

  47. [47]

    Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds,

    Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakoset al., “Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds,”arXiv, 2024

    2024

  48. [48]

    Free360: Layered gaussian splatting for unbounded 360-degree view synthesis from extremely sparse and unposed views,

    C. Bao, X. Zhang, Z. Yu, J. Shi, G. Zhang, S. Peng, and Z. Cui, “Free360: Layered gaussian splatting for unbounded 360-degree view synthesis from extremely sparse and unposed views,” inIEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 16 377– 16 387

    2025

  49. [49]

    Hgs-planner: Hierarchical planning framework for active scene reconstruction using 3d gaussian splatting,

    Z. Xu, R. Jin, K. Wu, Y . Zhao, Z. Zhang, J. Zhao, F. Gao, Z. Gan, and W. Ding, “Hgs-planner: Hierarchical planning framework for active scene reconstruction using 3d gaussian splatting,” inIEEE Intl. Conf. on Robotics and Automation (ICRA). IEEE, 2025, pp. 14 161–14 167

    2025

  50. [50]

    Activegs: Active scene reconstruction using gaussian splatting,

    L. Jin, X. Zhong, Y . Pan, J. Behley, C. Stachniss, and M. Popovi ´c, “Activegs: Active scene reconstruction using gaussian splatting,”IEEE Robotics and Automation Letters, 2025

    2025

  51. [51]

    Nextbest- path: Efficient 3d mapping of unseen environments,

    S. Li, A. Gu ´edon, C. Boittiaux, S. Chen, and V . Lepetit, “Nextbest- path: Efficient 3d mapping of unseen environments,”arXiv preprint arXiv:2502.05378, 2025

    Pith/arXiv arXiv 2025

  52. [52]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R¨adle, C. Rolland, L. Gustafsonet al., “Sam 2: Segment anything in images and videos,”arXiv, 2024

    2024

  53. [53]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, 2004

    2004

  54. [54]

    Monocular online reconstruction with enhanced detail preservation,

    S. Wu, Z. Lv, Y . Zhu, D. Frost, Z. Li, L.-Q. Yan, C. Ren, R. Newcombe, and Z. Dong, “Monocular online reconstruction with enhanced detail preservation,” inSIGGRAPH, 2025, pp. 1–11

    2025

  55. [55]

    A better way to construct the sunflower head,

    H. V ogel, “A better way to construct the sunflower head,”Mathematical biosciences, vol. 44, no. 3-4, pp. 179–189, 1979

    1979

  56. [56]

    The prize collecting traveling salesman problem,

    E. Balas, “The prize collecting traveling salesman problem,”Networks, vol. 19, no. 6, pp. 621–636, 1989

    1989

  57. [57]

    Finding the k shortest loopless paths in a network,

    J. Y . Yen, “Finding the k shortest loopless paths in a network,” management Science, vol. 17, no. 11, pp. 712–716, 1971

    1971

  58. [58]

    Hi-slam2: Geometry-aware gaussian slam for fast monocular scene reconstruction,

    W. Zhang, Q. Cheng, D. Skuddis, N. Zeller, D. Cremers, and N. Haala, “Hi-slam2: Geometry-aware gaussian slam for fast monocular scene reconstruction,”IEEE Trans. Robotics, vol. 41, pp. 6478–6493, 2025

    2025

  59. [59]

    Google scanned objects: A high- quality dataset of 3d scanned household items,

    L. Downs, A. Francis, N. Koenig, B. Kinman, R. Hickman, K. Reymann, T. B. McHugh, and V . Vanhoucke, “Google scanned objects: A high- quality dataset of 3d scanned household items,” inIEEE Intl. Conf. on Robotics and Automation (ICRA). IEEE, 2022, pp. 2553–2560

    2022