Pith. sign in

REVIEW 4 major objections 3 minor 5 cited by

This paper claims that a training-free temporal-spatial mask on the persistent state can contain catastrophic forgetting in streaming 3D reconstruction, showing only 1.33x error growth on long NRGBD sequences versus over 4x for the uniform-

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 06:29 UTC pith:YMHQFMQE

load-bearing objection Solid inference-time method for streaming 3D reconstruction, but the AND fusion in Eq. (10) undercuts the paper's complementary-signal story and the key threshold τ is never reported. the 4 major comments →

arxiv 2601.22615 v3 pith:YMHQFMQE submitted 2026-01-30 cs.CV

TTSA3R: Training-Free Temporal-Spatial Adaptive Persistent State for Streaming 3D Reconstruction

Zhijie Zheng , Xinhao Xiang , Jiawei Zhang This is my paper
classification cs.CV
keywords streaming 3D reconstructioncatastrophic forgettingpersistent statetraining-free adaptationtemporal-spatial update maskcross-attention gatinglong-sequence robustnessvideo depth estimation
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.

Streaming 3D reconstruction models keep a compact persistent state and update it every frame; the standard update replaces old geometry wholesale, so over long sequences errors accumulate and earlier structure is forgotten. The paper argues that the state should instead be updated selectively, token by token, based on two signals available at inference: how much each state token has changed across frames (temporal) and whether it is strongly attending to image regions whose features are changing (spatial). It proposes a training-free fusion of these signals into a mask, applies the mask as a weighted interpolation between the candidate new state and the old state, and reports that error growth is dramatically lower than uniform updates over 50–250 frames. A sympathetic reader would care because this is a recipe for long-horizon online reconstruction without retraining or extra memory.

Core claim

The central claim is that the uniform state-update rule in recurrent streaming reconstruction overwrites historically good geometry, causing drift and forgetting, and that this can be fixed without training by gating each state token with two complementary signals. The paper derives a per-token mask from temporal state evolution (normalized change magnitude between consecutive candidate states, sigmoid-gated) and from spatial context (cross-attention strength times frame-to-frame feature divergence, max-pooled over image tokens, sigmoid-gated). The two masks are multiplied and used in a masked interpolation: S_t = S̃_t ⊙ M_final + S_{t-1} ⊙ (1 − M_final). On NRGBD, as sequences extend from 5

What carries the argument

The load-bearing object is the fused temporal-spatial mask M_final = σ(Δ̂_t − τ) ⊙ σ(max_spat(A_t ⊙ D_t)). Here Δ̂_t is the per-token L2 norm of the candidate state change, normalized by its global average; A_t is the average absolute cross-attention from state tokens to image tokens across decoder layers; D_t is the cosine dissimilarity between consecutive frame features; and max_spat pools over image tokens. The mask multiplies the candidate update and its complement preserves the old state, so only tokens that are both temporally dynamic and spatially engaged are overwritten. All quantities come from the frozen pretrained decoder, making the mechanism training-free.

Load-bearing premise

The load-bearing premise is that small temporal change plus low cross-attention/feature-divergence reliably indicates already-correct geometry; if a token changes little only because views are redundant while the geometry is still wrong, the mask freezes the error.

What would settle it

Run the method on a long synthetic sequence whose first half consists of repeated frames of an intentionally corrupted reconstruction; if the mask freezes those tokens (low Δ̂ and low A⊙D) and the error remains while a uniform-update variant corrects it, the core premise fails. The paper should also report the threshold τ and a sweep over it to show the effect is not just parameter tuning.

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

If this is right

  • Long-video 3D reconstruction and online tracking can keep constant memory while resisting drift, because the mask is computed from internal signals of the frozen model.
  • Only tokens that are both temporally dynamic and spatially engaged get overwritten, so converged geometry is preserved indefinitely across long streams.
  • The method narrows the gap between streaming and full-attention pipelines on depth and pose benchmarks, e.g., TUM-dynamics ATE of 0.026 versus 0.046 for the baseline.
  • Since no training is involved, the same mask recipe can be transferred to other recurrent 3D architectures that expose state tokens and cross-attention.

Where Pith is reading between the lines

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

  • The paper's own limitation section concedes that performance degrades under severe occlusions or sparse observations, exactly where the spatial signal (cross-attention × feature divergence) becomes unreliable; this bounds the method to well-overlapped streams.
  • The paper never reports the threshold τ in the temporal gate and does not study sensitivity to it, so part of the reported gain could be tuning this one parameter rather than the mechanism itself.
  • The mask is only as good as the assumption that small temporal change plus low attention/divergence means correct geometry; on redundant-view segments of a wrong reconstruction, it would freeze errors instead of fixing them.
  • One testable extension is to monitor mask statistics as a signal for loop closure or scene novelty, since a sudden spike in tokens requiring update may indicate revisiting an area with new observations.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper presents TTSA3R, a training-free modification of the streaming 3D reconstruction model CUT3R. The method introduces two per-token gating signals: a Temporal Adaptive Update Module (TAUM) based on the normalized magnitude of change between consecutive candidate state tokens (Eqs. 5-6), and a Spatial Context Update Module (SCUM) based on the product of decoder cross-attention and inter-frame feature cosine dissimilarity (Eqs. 7-9). These masks are fused multiplicatively in Eq. (10) and used to interpolate between the previous persistent state and the new candidate state (Eq. 11). The authors evaluate on video depth estimation, camera pose estimation, and 3D reconstruction, and claim substantially improved robustness on long sequences, e.g., only a 1.33x error increase on NRGBD from 50 to 250 frames versus over 4x for CUT3R.

Significance. If the long-sequence claims hold, the paper would offer a simple, training-free way to mitigate catastrophic forgetting in recurrent streaming 3D reconstruction, with broad applicability to depth, pose, and geometry tasks. The method is clearly described, the code is promised, and the ablations are informative. However, the current evidentiary basis for the central robustness claim is incomplete, and one key design choice appears to contradict the paper's own motivation.

major comments (4)
  1. [Eq. (10) with Sec. 3.3] The multiplicative fusion Mfinal = Mtemp ⊙ Mspat blocks the exact case used to motivate SCUM. Section 3.3 states that a token 'may change minimally across frames because prior views lacked coverage, yet current views provide new spatial information.' For such a token, Mtemp ≈ 0 (small Δ̂t in Eq. (6)) while Mspat ≈ 1 (new coverage implies high attention and high feature divergence), so Eq. (10) gives Mfinal ≈ 0 and the update is suppressed. Thus SCUM cannot act as a complement to TAUM in the motivating scenario; the only tokens updated are those with both large temporal change and large spatial signal. The paper needs either a different fusion rule (e.g., a maximum/OR form or a learned combination) or an explicit argument for why the AND gate is nevertheless appropriate. The Limitations paragraph in Sec. 5, which concedes that correspondence signals become unreliable under sparse observat
  2. [Eq. (6) and Sec. 4 (Implementation Details)] The temporal threshold τ is never reported, and no sensitivity study is provided. Since Mtemp = σ(Δ̂t − τ), τ directly controls how often tokens can update; without its value or a sweep, the reader cannot tell whether the reported long-sequence gains come from a meaningfully adaptive gate or from a nearly always frozen state. Because TTSA3R is described as training-free, it is also important to state whether τ was chosen from validation performance on the evaluated benchmarks. This omission is load-bearing for the central claim and should be addressed with the exact value, the selection procedure, and a sensitivity analysis.
  3. [Sec. 4.3, Figs. 4-6] The central long-sequence claim—'only 1.33x error increase compared to over 4x degradation'—is supported by curves without error bars, on a single dataset (NRGBD) for the 3D reconstruction result, and with no statement of the number of test sequences or independent runs. Please report exact numeric values at 50 vs. 250 frames, standard deviations or per-sequence statistics, and analogous quantitative tables for the long-sequence depth (Fig. 4) and pose (Fig. 5) evaluations. Without this, the robustness claim is difficult to assess and cannot be compared against the short-sequence tables in the paper.
  4. [Eq. (5)] The temporal evolution term uses S̃_t − S̃_{t−1}, but the algorithm as described stores S_t via Eq. (11) and never specifies that the previous candidate state S̃_{t−1} is retained. If S̃_{t−1} is not kept, Eq. (5) is not computable online; if it is kept, that extra memory should be stated in the implementation details and accounted for in the efficiency analysis. In addition, the dimensionality of D_t in Eq. (7) and the exact alignment of A_t ⊙ D_t in Eq. (9) are not defined precisely, which hampers reproduction.
minor comments (3)
  1. [Sec. 4.3] 'reformative quality' appears to be a typo; likely 'refined quality' or 'improved quality'.
  2. [Sec. 4, Implementation Details] The statement 'Our codes are available' should specify the exact code version or commit and list the main dependencies, especially because the method requires access to internal cross-attention maps from a pretrained CUT3R model.
  3. [Eqs. (7)-(9)] The notation D_t is used both for decoder features in Eq. (2) and for cosine dissimilarity in Eq. (7). Using distinct symbols would avoid confusion.

Circularity Check

0 steps flagged

No significant circularity: the method is an empirical inference-time gating scheme on a frozen base model, with no fitted parameter renamed as a prediction.

full rationale

The paper's derivation chain is an inference-time masking scheme applied to a frozen pretrained CUT3R. The temporal mask is a sigmoid-gated normalized norm of candidate-state change; the spatial mask is a sigmoid-gated max-pooled product of cross-attention and cosine dissimilarity; the final state update is a convex combination controlled by their element-wise product. None of these quantities is fitted to the benchmark metrics, and no parameter is learned or optimized against the reported test sets. The only free scalar τ is a hand-set threshold and is not reported, which is a reproducibility/tuning ambiguity rather than a demonstrable circular reduction. The long-sequence robustness claim (1.33x vs over 4x degradation on NRGBD) is an empirical comparison against baselines, not a result derived from the defining equations. I also considered the potential self-citations in references [55] and [59], since their author names overlap with the present authors, but those citations appear only in related-work and benchmark contexts and are not load-bearing. The AND-fusion inconsistency between SCUM's motivating example and Eq. (10), and the Limitations' caveat about unreliable correspondence under sparse observations, are correctness/design concerns, not cases where a prediction is equivalent to its input by construction. Therefore no specific circular step can be quoted, and the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

The method is a heuristic gating rule over CUT3R's internal state. It introduces one free hyperparameter (τ, unreported). It relies on three interpretational assumptions about what internal signals mean. It introduces no new entities or fitted parameters in the sense of learned weights.

free parameters (1)
  • τ (temporal mask threshold) = not reported
    Threshold in sigmoid gating of the temporal mask (Eq. 6). Controls what counts as significant state change. Its value is not given in the paper, so the results depend on an unreported choice.
axioms (3)
  • domain assumption Cross-attention between state tokens and image features reflects alignment confidence and update necessity
    Invoked in Sec. 3.3, Eq. (8): the method assumes high cross-attention to changing regions indicates active refinement, while stable regions should be preserved.
  • domain assumption Cosine dissimilarity between consecutive frame features detects meaningful scene change
    Invoked in Sec. 3.3, Eq. (7): feature divergence is treated as a proxy for geometric change requiring update.
  • domain assumption Per-token temporal state-change magnitude correlates with convergence or refinement need
    Invoked in Sec. 3.2, Eq. (5)-(6): tokens that change little are assumed to be stable geometry worth preserving, and tokens that change much are assumed to need aggressive updates.

pith-pipeline@v1.3.0-alltime-deepseek · 11955 in / 9470 out tokens · 95092 ms · 2026-08-03T06:29:13.925514+00:00 · methodology

0 comments
read the original abstract

Streaming recurrent models enable efficient 3D reconstruction by maintaining persistent state representations. However, they suffer from catastrophic forgetting over long sequences due to balancing historical information with new observations. Recent methods alleviate this by deriving adaptive signals from the attention perspective, but they operate on single dimensions without considering temporal and spatial consistency. To this end, we propose a training-free framework termed TTSA3R that leverages both temporal state evolution and spatial observation quality for adaptive state updates in 3D reconstruction. In particular, we devise a Temporal Adaptive Update Module that regulates update magnitude by analyzing temporal state evolution patterns. Then, a Spatial Contextual Update Module is introduced to localize spatial regions that require updates through observation-state alignment and scene dynamics. These complementary signals are finally fused to determine the state updating strategies. Extensive experiments show that TTSA3R achieves competitive performance on standard short-sequence benchmarks and provides substantially stronger robustness on extended sequences. On NRGBD, as sequences extend from 50 to 250 frames, TTSA3R exhibits only a 1.33x error increase, compared with over 4x degradation for CUT3R. This highlights the practical value of temporal-spatial adaptive updates for long-term reconstruction stability. Our code is available at https://github.com/anonus2357/ttsa3r.

Figures

Figures reproduced from arXiv: 2601.22615 by Jiawei Zhang, Xinhao Xiang, Zhijie Zheng.

Figure 1
Figure 1. Figure 1: Catastrophic forgetting in streaming 3D reconstruction. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: , it consists of two main components: Temporal Adaptive Update Module and Spatial Context Update Module. These two modules provide complementary signals that are integrated for fine-grained state updates [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of Spatial Context Update Module. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Video depth estimation (long sequences) using metric depth accuracy on Bonn [35] dataset. Bonn [35] dataset and approaches the performance of full-attention pipelines based on metric scale, narrowing the gap between online and offline methods [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Camera pose estimation (long sequences) on TUM-dynamics [42] (left) and ScanNet [14] (right) datasets. 4.3 3D Reconstruction We further evaluate 3D reconstruction on NRGBD [2] dataset, as shown in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 3D reconstruction (long sequences) on NRGBD [2] dataset [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of 3D reconstruction. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of predicted camera trajectories [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of inference efficiency. We compare GPU memory used and FPS of different methods using 512×144 image resolution on KITTI [19] dataset. Inference Efficiency Analysis. We further evaluate computational efficiency by measuring inference speed and GPU memory consumption on video depth estimation. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] 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. Mem3R: Streaming 3D Reconstruction with Hybrid Memory via Test-Time Training

    cs.CV 2026-04 unverdicted novelty 7.0

    Mem3R achieves better long-sequence 3D reconstruction by decoupling tracking and mapping with a hybrid memory of TTT-updated MLP and explicit tokens, reducing model size and trajectory errors.

  2. Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers

    cs.CV 2026-05 unverdicted novelty 6.0

    A two-stage diversity-plus-entropy token selection framework speeds up visual geometry transformers by over 85% on 500-image scenes while preserving baseline accuracy.

  3. Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction

    cs.CV 2026-05 unverdicted novelty 6.0

    A closed-form scalar frame-level gate α_t derived from internal feature changes extends effective memory in recurrent 3D reconstruction and improves accuracy on long sequences up to 4541 frames.

  4. Attention Itself Could Retrieve.RetrieveVGGT: Training-Free Long Context Streaming 3D Reconstruction via Query-Key Similarity Retrieval

    cs.CV 2026-05 unverdicted novelty 6.0

    RetrieveVGGT enables constant-memory long-context streaming 3D reconstruction by retrieving relevant frames via query-key similarities in VGGT's first attention layer, outperforming StreamVGGT and others.

  5. ReCal3R: Reliability-Calibrated Learning Rates for Streaming 3D Reconstruction

    cs.CV 2026-07 conditional novelty 5.0

    Calibrating per-token learning rates by estimated state-token reliability, not only observation alignment, stabilizes long-stream recurrent 3D reconstruction with large ATE gains and similar cost.

Reference graph

Works this paper leans on

62 extracted references · 10 linked inside Pith · cited by 5 Pith papers

  1. [1]

    Communications of the ACM54(10), 105–112 (2011)

    Agarwal, S., Furukawa, Y., Snavely, N., Simon, I., Curless, B., Seitz, S.M., Szeliski, R.: Building rome in a day. Communications of the ACM54(10), 105–112 (2011)

    2011

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Azinović, D., Martin-Brualla, R., Goldman, D.B., Nießner, M., Thies, J.: Neural rgb-d surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6290–6301 (2022)

    2022

  3. [3]

    In: European conference on computer vision

    Butler, D.J., Wulff, J., Stanley, G.B., Black, M.J.: A naturalistic open source movie for optical flow evaluation. In: European conference on computer vision. pp. 611–

  4. [4]

    IEEE Transactions on robotics32(6), 1309– 1332 (2017)

    Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid, I., Leonard, J.J.: Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics32(6), 1309– 1332 (2017)

    2017

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Charatan,D.,Li,S.L.,Tagliasacchi,A.,Sitzmann,V.:pixelsplat:3dgaussiansplats from image pairs for scalable generalizable 3d reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 19457– 19467 (2024)

    2024

  6. [6]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Chen, H., Gu, J., Chen, A., Tian, W., Tu, Z., Liu, L., Su, H.: Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 2416–2425 (2023)

    2023

  7. [7]

    In: Pro- ceedings of the IEEE/CVF international conference on computer vision

    Chen, R., Han, S., Xu, J., Su, H.: Point-based multi-view stereo network. In: Pro- ceedings of the IEEE/CVF international conference on computer vision. pp. 1538– 1547 (2019)

    2019

  8. [8]

    arXiv preprint arXiv:2503.24391 (2025)

    Chen, X., Chen, Y., Xiu, Y., Geiger, A., Chen, A.: Easi3r: Estimating disentangled motion from dust3r without training. arXiv preprint arXiv:2503.24391 (2025)

    arXiv 2025

  9. [9]

    arXiv preprint arXiv:2509.26645 (2025)

    Chen, X., Chen, Y., Xiu, Y., Geiger, A., Chen, A.: Ttt3r: 3d reconstruction as test-time training. arXiv preprint arXiv:2509.26645 (2025)

    Pith/arXiv arXiv 2025

  10. [10]

    arXiv preprint arXiv:2506.02112 (2025)

    Chen,X.,Xia,T.,Xu,S.,Yang,J.,Chai,J.,Cheng,Z.:Sab3r:Semantic-augmented backbone in 3d reconstruction. arXiv preprint arXiv:2506.02112 (2025)

    Pith/arXiv arXiv 2025

  11. [11]

    arXiv preprint arXiv:2510.06219 (2025)

    Chen, Y., Chen, X., Xue, Y., Chen, A., Xiu, Y., Pons-Moll, G.: Human3r: Everyone everywhere all at once. arXiv preprint arXiv:2510.06219 (2025)

    arXiv 2025

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Chen, Z., Qin, M., Yuan, T., Liu, Z., Zhao, H.: Long3r: Long sequence streaming 3d reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5273–5284 (2025)

    2025

  13. [13]

    In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition

    Cui, H., Gao, X., Shen, S., Hu, Z.: Hsfm: Hybrid structure-from-motion. In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition. pp. 1212–1221 (2017)

    2017

  14. [14]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5828–5839 (2017)

    2017

  15. [15]

    In: Proceedings Ninth IEEE International Conference on Computer Vision

    Davison: Real-time simultaneous localisation and mapping with a single camera. In: Proceedings Ninth IEEE International Conference on Computer Vision. pp. 1403–1410. IEEE (2003)

    2003

  16. [16]

    IEEE transactions on pattern analysis and machine intelligence 29(6), 1052–1067 (2007)

    Davison, A.J., Reid, I.D., Molton, N.D., Stasse, O.: Monoslam: Real-time sin- gle camera slam. IEEE transactions on pattern analysis and machine intelligence 29(6), 1052–1067 (2007)

    2007

  17. [17]

    arXiv preprint arXiv:2010.11929 (2020) 16 Zheng et al

    Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. arXiv preprint arXiv:2010.11929 (2020) 16 Zheng et al

    Pith/arXiv arXiv 2010

  18. [18]

    IEEE robotics & automation magazine13(2), 99–110 (2006)

    Durrant-Whyte, H., Bailey, T.: Simultaneous localization and mapping: part i. IEEE robotics & automation magazine13(2), 99–110 (2006)

    2006

  19. [19]

    The international journal of robotics research32(11), 1231–1237 (2013)

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. The international journal of robotics research32(11), 1231–1237 (2013)

    2013

  20. [20]

    In: European Conference on Computer Vision

    Guo, Y., Kumar, A., Zhao, C., Wang, R., Huang, X., Ren, L.: Sup-nerf: A stream- lined unification of pose estimation and nerf for monocular 3d object reconstruc- tion. In: European Conference on Computer Vision. pp. 37–53. Springer (2024)

    2024

  21. [21]

    In: 2007 IEEE Conference on Computer Vision and Pattern Recognition

    Habbecke, M., Kobbelt, L.: A surface-growing approach to multi-view stereo recon- struction. In: 2007 IEEE Conference on Computer Vision and Pattern Recognition. pp. 1–8. IEEE (2007)

    2007

  22. [22]

    ACM Transactions on Graphics (TOG) 38(1), 1–17 (2018)

    Hepp, B., Nießner, M., Hilliges, O.: Plan3d: Viewpoint and trajectory optimization for aerial multi-view stereo reconstruction. ACM Transactions on Graphics (TOG) 38(1), 1–17 (2018)

    2018

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Huang, H., Li, L., Cheng, H., Yeung, S.K.: Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21584–21593 (2024)

    2024

  24. [24]

    ACM Transactions on Graphics42(4), 1–14 (2023)

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4), 1–14 (2023)

    2023

  25. [25]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Kopf, J., Rong, X., Huang, J.B.: Robust consistent video depth estimation. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 1611–1621 (2021)

    2021

  26. [26]

    arXiv preprint arXiv:2508.10893 (2025)

    Lan,Y.,Luo,Y.,Hong,F.,Zhou,S.,Chen,H.,Lyu,Z.,Yang,S.,Dai,B.,Loy,C.C., Pan, X.: Stream3r: Scalable sequential 3d reconstruction with causal transformer. arXiv preprint arXiv:2508.10893 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

    In: European Conference on Computer Vision

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: European Conference on Computer Vision. pp. 71–91. Springer (2024)

    2024

  28. [28]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA)

    Li, F., Hu, P., Song, Q., Huang, R.: Incremental 3d reconstruction through a hybrid explicit-and-implicit representation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). pp. 15121–15127. IEEE (2024)

    2024

  29. [29]

    arXiv preprint arXiv:2509.05296 (2025)

    Li, Z., Zhou, J., Wang, Y., Guo, H., Chang, W., Zhou, Y., Zhu, H., Chen, J., Shen, C., He, T.: Wint3r: Window-based streaming reconstruction with camera token pool. arXiv preprint arXiv:2509.05296 (2025)

    Pith/arXiv arXiv 2025

  30. [30]

    In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

    Lindenberger, P., Sarlin, P.E., Larsson, V., Pollefeys, M.: Pixel-perfect structure- from-motion with featuremetric refinement. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 5987–5997 (2021)

    2021

  31. [31]

    In: European Conference on Computer Vision

    Liu, S., Gao, Y., Zhang, T., Pautrat, R., Schönberger, J.L., Larsson, V., Pollefeys, M.: Robust incremental structure-from-motion with hybrid features. In: European Conference on Computer Vision. pp. 249–269. Springer (2024)

    2024

  32. [32]

    Commu- nications of the ACM65(1), 99–106 (2021)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021)

    2021

  33. [33]

    In: 2011 international conference on computer vision

    Newcombe, R.A., Lovegrove, S.J., Davison, A.J.: Dtam: Dense tracking and map- ping in real-time. In: 2011 international conference on computer vision. pp. 2320–

    2011

  34. [34]

    In: 2007 IEEE 11th International Conference on Computer Vision

    Ni, K., Steedly, D., Dellaert, F.: Out-of-core bundle adjustment for large-scale 3d reconstruction. In: 2007 IEEE 11th International Conference on Computer Vision. pp. 1–8. IEEE (2007) TTSA3R 17

    2007

  35. [35]

    In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Palazzolo, E., Behley, J., Lottes, P., Giguere, P., Stachniss, C.: Refusion: 3d recon- struction in dynamic environments for rgb-d cameras exploiting residuals. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 7855–7862. IEEE (2019)

    2019

  36. [36]

    In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops

    Pinard, C., Chevalley, L., Manzanera, A., Filliat, D.: Learning structure-from- motion from motion. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. pp. 0–0 (2018)

    2018

  37. [37]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104–4113 (2016)

    2016

  38. [38]

    In: 2014 2nd International Conference on 3D Vision

    Shah, R., Deshpande, A., Narayanan, P.: Multistage sfm: Revisiting incremental structure from motion. In: 2014 2nd International Conference on 3D Vision. vol. 1, pp. 417–424. IEEE (2014)

    2014

  39. [39]

    arXiv preprint arXiv:2512.03939 (2025)

    Shen, G., Deng, T., Qin, X., Wang, N., Wang, J., Wang, Y., Chen, Y., Wang, H., Wang, J.: Mut3r: Motion-aware updating transformer for dynamic 3d reconstruc- tion. arXiv preprint arXiv:2512.03939 (2025)

    arXiv 2025

  40. [40]

    arXiv preprint arXiv:2509.02560 (2025)

    Shen, Y., Zhang, Z., Qu, Y., Zheng, X., Ji, J., Zhang, S., Cao, L.: Fastvggt: Training-free acceleration of visual geometry transformer. arXiv preprint arXiv:2509.02560 (2025)

    Pith/arXiv arXiv 2025

  41. [41]

    In: ACM siggraph 2006 papers, pp

    Snavely, N., Seitz, S.M., Szeliski, R.: Photo tourism: exploring photo collections in 3d. In: ACM siggraph 2006 papers, pp. 835–846 (2006)

    2006

  42. [42]

    In: 2012 IEEE/RSJ international conference on intelligent robots and systems

    Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573–580. IEEE (2012)

    2012

  43. [43]

    In: Proceed- ings of the Computer Vision and Pattern Recognition Conference

    Tang, J., Gao, Y., Yang, D., Yan, L., Yue, Y., Yang, Y.: Dronesplat: 3d gaussian splatting for robust 3d reconstruction from in-the-wild drone imagery. In: Proceed- ings of the Computer Vision and Pattern Recognition Conference. pp. 833–843 (2025)

    2025

  44. [44]

    IEEE transactions on pattern analysis and machine intelligence30(5), 878–892 (2008)

    Torresani, L., Hertzmann, A., Bregler, C.: Nonrigid structure-from-motion: Esti- mating shape and motion with hierarchical priors. IEEE transactions on pattern analysis and machine intelligence30(5), 878–892 (2008)

    2008

  45. [45]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  46. [46]

    arXiv preprint arXiv:2509.07120 (2025)

    Wang, C.S.B., Schmidt, C., Piekenbrinck, J., Leibe, B.: Faster vggt with block- sparse global attention. arXiv preprint arXiv:2509.07120 (2025)

    Pith/arXiv arXiv 2025

  47. [47]

    arXiv preprint arXiv:2408.16061 (2024)

    Wang, H., Agapito, L.: 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061 (2024)

    Pith/arXiv arXiv 2024

  48. [48]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025)

    2025

  49. [49]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, Q., Zhang, Y., Holynski, A., Efros, A.A., Kanazawa, A.: Continuous 3d perception model with persistent state. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10510–10522 (2025)

    2025

  50. [50]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20697–20709 (2024)

    2024

  51. [51]

    Advances in Neural Information Process- ing Systems35, 3502–3516 (2022) 18 Zheng et al

    Weinzaepfel, P., Leroy, V., Lucas, T., Brégier, R., Cabon, Y., Arora, V., Antsfeld, L., Chidlovskii, B., Csurka, G., Revaud, J.: Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. Advances in Neural Information Process- ing Systems35, 3502–3516 (2022) 18 Zheng et al

    2022

  52. [52]

    In: 2013 Interna- tional Conference on 3D Vision-3DV 2013

    Wu, C.: Towards linear-time incremental structure from motion. In: 2013 Interna- tional Conference on 3D Vision-3DV 2013. pp. 127–134. IEEE (2013)

    2013

  53. [53]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu, R., Mildenhall, B., Henzler, P., Park, K., Gao, R., Watson, D., Srinivasan, P.P., Verbin, D., Barron, J.T., Poole, B., et al.: Reconfusion: 3d reconstruction with diffusion priors. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21551–21561 (2024)

    2024

  54. [54]

    arXiv preprint arXiv:2507.02863 (2025)

    Wu, Y., Zheng, W., Zhou, J., Lu, J.: Point3r: Streaming 3d reconstruction with explicit spatial pointer memory. arXiv preprint arXiv:2507.02863 (2025)

    arXiv 2025

  55. [55]

    In: 2025 IEEE International Conference on Robotics and Automation (ICRA)

    Xiang, H., Li, X., Cheng, K., Lai, X., Zhang, W., Liao, Z., Zeng, L., Liu, X.: Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction. In: 2025 IEEE International Conference on Robotics and Automation (ICRA). pp. 2686–2693. IEEE (2025)

    2025

  56. [56]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yang, J., Sax, A., Liang, K.J., Henaff, M., Tang, H., Cao, A., Chai, J., Meier, F., Feiszli, M.: Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 21924–21935 (2025)

    2025

  57. [57]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025)

    Yuan, Y., Shen, Q., Wang, S., Yang, X., Wang, X.: Test3r: Learning to reconstruct 3d at test time. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025)

    2025

  58. [58]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Yuan, Z., Huang, H., Xiong, Z., Wang, D., Yang, G.: Robust and efficient 3d gaus- sian splatting for urban scene reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 26209–26219 (2025)

    2025

  59. [59]

    arXiv preprint arXiv:2410.03825 (2024)

    Zhang, J., Herrmann, C., Hur, J., Jampani, V., Darrell, T., Cole, F., Sun, D., Yang, M.H.: Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825 (2024)

    Pith/arXiv arXiv 2024

  60. [60]

    In: European Conference on Computer Vision

    Zhang, K., Bi, S., Tan, H., Xiangli, Y., Zhao, N., Sunkavalli, K., Xu, Z.: Gs-lrm: Large reconstruction model for 3d gaussian splatting. In: European Conference on Computer Vision. pp. 1–19. Springer (2024)

    2024

  61. [61]

    In: European Conference on Computer Vision

    Zhang, Z., Cole, F., Li, Z., Rubinstein, M., Snavely, N., Freeman, W.T.: Structure and motion from casual videos. In: European Conference on Computer Vision. pp. 20–37. Springer (2022)

    2022

  62. [62]

    arXiv preprint arXiv:2507.11539 (2025)

    Zhuo, D., Zheng, W., Guo, J., Wu, Y., Zhou, J., Lu, J.: Streaming 4d visual geometry transformer. arXiv preprint arXiv:2507.11539 (2025)

    Pith/arXiv arXiv 2025