Pith. sign in

REVIEW 2 major objections 5 minor 49 references

Global attention in feed-forward 3D reconstruction transformers is highly heterogeneous and dynamic across heads; SAF3R exploits that with training-free head-specific sparse kernels to keep pose and reconstruction quality while cutting end-

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 · grok-4.5

2026-07-12 01:11 UTC pith:IXE7VVKS

load-bearing objection Solid training-free systems paper: F3R head taxonomy plus matched sparse kernels give real speedups with mostly preserved pose/recon; DINO Top-K is the known soft hinge, not a hidden break. the 2 major comments →

arxiv 2607.03612 v1 pith:IXE7VVKS submitted 2026-07-03 cs.CV cs.LG

SAF3R: Dynamic Sparse Attention for Feed-Forward 3D Reconstruction Transformers

Jianing Deng , Yuanzhe Li , Jialu Wang , Song Wang , Tianlong Chen , Huanrui Yang , Jingtong Hu This is my paper
classification cs.CV cs.LG
keywords feed-forward 3D reconstructionsparse attentiontransformerscamera pose estimationpoint cloud reconstructiondynamic sparsityhead-wise profiling
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.

Feed-forward 3D reconstruction transformers jointly estimate camera poses and dense geometry from many images by letting tokens from every view attend to every other view. That global attention is quadratic in the number of tokens and quickly becomes the bottleneck on long sequences. Prior speedups use token merging or a single fixed sparse pattern shared across the model, which underuses the fact that different layers and heads behave very differently. This paper first measures those patterns on several large models and shows they are extremely sparse, change with the input, and fall into a few functional types. It then builds SAF3R: a training-free pipeline that assigns each head one of four tailored sparse kernels, chooses the assignment offline on a small calibration set by progressive substitution that never raises approximation error, and adapts the dynamic kernels online with cheap probes and cached indices. The result is high sparsity with camera-pose and point-cloud quality close to full attention, and large measured end-to-end speedups on long inputs.

Core claim

Global attention heads in F3R transformers are not uniform: they fall into position, vertical-line, correspondence, and scanning types whose sparsity and input-dependence differ, so a single static or model-level sparsity pattern wastes either accuracy or compute. Assigning each head a matching sparse kernel via offline profiling and online adaptation recovers nearly full-attention pose and reconstruction quality at far higher sparsity than prior efficient methods.

What carries the argument

SAF3R: a training-free two-stage dynamic sparse attention framework that combines four head-type kernels (static, query-probe Top-K, DINO Top-K, uniform sampling), offline progressive-substitution profiling that minimizes NMSE under a cost budget, and online O(N) adaptation that estimates Top-K keys from a mean-query probe or cached DINO similarities.

Load-bearing premise

That DINO patch cosine similarity is a good enough stand-in for the keys that correspondence heads actually attend to, so a small per-frame Top-K recovers the geometric signal needed for pose and reconstruction.

What would settle it

On scenes where DINO features fail to match true correspondences (textureless regions, repeated patterns, strong illumination change, or large viewpoint gaps), measure whether small-K DINO Top-K kernels still keep AUC@3 and Chamfer distance near the full-attention baseline; if accuracy collapses while uniform sampling at similar cost does not, the proxy assumption fails.

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

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

2 major / 5 minor

Summary. The paper analyzes cross-view global attention in feed-forward 3D reconstruction (F3R) transformers (VGGT, π3, MapAnything, DA3) and reports that patterns are heterogeneous across layers/heads and often highly sparse and input-dependent. It categorizes heads into position, vertical-line, correspondence, and scanning types and designs matching sparse kernels (static, Query-Probe Top-K, DINO Top-K, uniform sampling). SAF3R is a training-free two-stage system: offline progressive substitution of head configs that minimizes NMSE to full attention under a compute budget on a small calibration set (Eqs. 1–2), plus online adaptation (O(N) mean-query probe for vertical-line heads; cached DINO cosine Top-K for correspondence heads). Experiments on sparse and dense multi-view benchmarks claim pose and reconstruction quality close to full attention and better than several token-merging/sparse baselines, with end-to-end speedups up to ~7× on long sequences (Tables 2–4, Fig. 6).

Significance. If the results hold, SAF3R is a practical, training-free acceleration path for a growing class of multi-view geometric transformers whose global attention is the main long-sequence bottleneck. Strengths include multi-model head analysis, head-specific kernels rather than model-level static sparsity, an explicit offline NMSE search, multi-dataset pose/reconstruction tables under sparse and dense settings, efficiency curves, and ablations on profiling and DINO Top-K. Code is promised. The contribution is solid systems/engineering work for the F3R community rather than a new geometric principle; its value depends on how robust the DINO proxy and calibration transfer remain outside the reported regimes.

major comments (2)
  1. Sec. 4.1 (DINO Top-K), Sec. 5.3, Limitations, and Fig. 8: the central quality-at-high-sparsity claim rests on DINO cosine similarity as a proxy for correspondence-head keys. The paper itself shows failure modes (textureless regions, repeats, illumination/viewpoint change) that force larger K and erode sparsity; HiRoom and Oxford Spires illustrate the trade-off. The main tables report a single operating K range without a systematic sensitivity or worst-case protocol across all datasets. A clearer default-K policy, failure-rate summary, and/or the hybrid Top-K+stride fallback suggested in Sec. 5.3 would make the claim load-bearing rather than conditional on favorable scenes.
  2. Sec. 4.2–5.1 and Eqs. (1)–(2): offline profiling uses 10 ETH3D training sequences and NMSE as the sole selection criterion under a uniform-stride baseline (σ≈0.85). Transfer of these head assignments to Co3D, RealEstate10K, ScanNet, HiRoom, DTU, etc., is assumed but not stress-tested (e.g., re-profile on another domain or report NMSE vs. downstream AUC/CD correlation). Because the method is training-free and head configs are frozen after calibration, evidence that the proxy and search generalize is necessary for the multi-benchmark claims in Tables 2–4.
minor comments (5)
  1. Table 1 and Sec. 2.3: AVGGT is re-implemented without public code; a short appendix note on fidelity to the paper would strengthen the baseline comparison.
  2. Fig. 7 and appendix: profiled kernel maps are useful; stating the fraction of heads per type and the resulting FLOPs/sparsity per model in the main text would make the heterogeneity claim more quantitative.
  3. Sec. 4.3: block-wise vs. token-wise DINO Top-K is mentioned but not quantified in the main efficiency or accuracy tables; a small ablation would clarify the accuracy–overhead trade-off.
  4. Dense DA3 rows in Table 4 hit OOM for several methods; clarifying memory accounting (including Top-K index cache) would help interpret the speedup claims at 700 frames.
  5. Minor presentation: consistent naming of π3 / Aπ3 / Fastπ3 across tables; ensure appendix Figs. 9–11 are referenced from the main analysis section.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper whose sparse kernels and offline search minimize approximation error against full attention, then evaluate on held-out pose/reconstruction benchmarks.

full rationale

SAF3R is a training-free engineering framework, not a first-principles derivation. Head types are observational categories from attention-map analysis (Sec. 3); kernels are designed to match those patterns (static, Query-Probe Top-K, DINO Top-K, uniform sampling) and reduce complexity from O(N^{2}) to O(N). Offline profiling (Eqs. 1–2) chooses, for each head, the candidate configuration that minimizes NMSE to the full-attention output under a FLOPs budget no higher than a uniform-stride baseline; the search does not define the target metrics (AUC, ATE, CD, F1) in terms of the fit. Online adaptation (mean-query probe for vertical-line heads; cached DINO cosine Top-K for correspondence heads) is an efficiency reformulation, not a tautology. Success is measured on multi-dataset camera-pose and point-cloud benchmarks against full-attention baselines and prior efficient methods (Tables 2–4, Fig. 6). The DINO-proxy assumption is an empirical approximation whose failure modes the paper itself documents; it is not circular. No self-definitional loop, fitted-input-as-prediction, load-bearing self-citation uniqueness claim, or renaming of a known result appears. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim rests on empirical modeling choices rather than new physical entities: a four-type head taxonomy, four sparse kernels, a calibration-driven greedy assignment, and free knobs (baseline sparsity, K, calibration set). Domain assumptions include that full multi-head attention can be approximated head-wise with O(N) patterns and that DINO features proxy correspondence. No new particles or forces; invented entities are methodological constructs.

free parameters (4)
  • baseline uniform sparsity ratio σ
    Offline search starts from uniform sampling with σ=0.85 (Sec. 5.1); this sets the compute budget and lower bound on sparsity for all subsequent head substitutions.
  • DINO Top-K per-frame K
    K is set between 16 and 128 (and varied to 256 in analysis) depending on scene; accuracy–sparsity trade-off is sensitive to K, especially outdoors (Sec. 5.1, 5.3, Fig. 8).
  • calibration dataset size and source
    Head configs are chosen using 10 ETH3D training sequences; generalization of the profiled assignment depends on this small hand-chosen set (Sec. 5.1).
  • candidate kernel settings in Ch (K values, static patterns, strides)
    The progressive substitution only chooses among author-defined candidates with equal or higher sparsity than the baseline; the discrete menu is a design choice that shapes the final heterogeneous config (Sec. 4.2).
axioms (5)
  • domain assumption Global attention in F3R transformers is sufficiently sparse and structured that O(N) head-specific patterns can approximate full attention without large task error.
    Motivated by Sec. 3 observations and used as the premise for all kernels and the speedup claim.
  • ad hoc to paper Attention heads can be usefully partitioned into position, vertical-line, correspondence, and scanning types with distinct sparse kernels.
    Taxonomy introduced in Observation #2 (Sec. 3) and operationalized in Sec. 4.1; not a standard theorem, but an empirical categorization.
  • domain assumption DINO patch features provide adequate cross-view correspondence estimates for correspondence heads.
    Stated in Sec. 4.1 and supported by appendix overlap figures; explicitly fragile in Limitations.
  • standard math Column-wise average of pre-softmax scores equals similarity of keys to the mean query, so Query-Probe Top-K ranking is exact for that criterion at O(N) cost.
    Linearity argument in Sec. 4.3; standard linear algebra, not an empirical fit.
  • ad hoc to paper NMSE between full and sparse attention outputs on a small calibration set is a valid proxy for downstream pose/reconstruction fidelity when selecting head configs.
    Eq. (1)–(2) define the offline objective; task metrics are only checked after assignment.
invented entities (2)
  • Four F3R global-attention head types (position, vertical-line, correspondence, scanning) no independent evidence
    purpose: Organize heterogeneous attention so each head can be assigned a specialized sparse kernel.
    Descriptive categories from visualization; independent evidence is only the paper’s attention maps and ablations, not an external standard taxonomy.
  • SAF3R offline progressive substitution profiler + online pattern adaptation independent evidence
    purpose: Assign and execute head-wise sparse configs without training.
    Methodological system; falsifiable via third-party reimplementation and benchmarks, but not an external physical entity.

pith-pipeline@v1.1.0-grok45 · 25768 in / 3656 out tokens · 31066 ms · 2026-07-12T01:11:38.814139+00:00 · methodology

0 comments
read the original abstract

Feed-forward 3D reconstruction (F3R) transformers have recently achieved remarkable success. However, scaling them to long image sequences remains challenging, as the quadratic complexity of cross-view global attention quickly becomes the dominant computational bottleneck. While recent efforts attempt to improve efficiency through compressed or sparse attention, they fail to fully exploit the inherent sparsity and dynamic behavior of global attention. In this work, we present a comprehensive analysis of global attention across multiple F3R transformers and reveal that attention patterns are highly heterogeneous, dynamic, and extremely sparse across layers and attention heads. Motivated by these findings, we propose SAF3R, a training-free dynamic sparse attention framework tailored to F3R transformers. SAF3R integrates tailored sparse attention mechanisms with offline head profiling and an efficient online adaptation strategy to match input-dependent attention behaviors. Extensive experiments demonstrate that SAF3R achieves high sparsity ratios while preserving camera pose estimation and 3D reconstruction quality, translating into substantial end-to-end speedup on F3R transformers compared to existing methods. Code is available at https://github.com/jndeng/SAF3R

Figures

Figures reproduced from arXiv: 2607.03612 by Huanrui Yang, Jialu Wang, Jianing Deng, Jingtong Hu, Song Wang, Tianlong Chen, Yuanzhe Li.

Figure 1
Figure 1. Figure 1: Left: SAF3R accelerates feed-forward 3D reconstruction (F3R) models while preserving reconstruction quality. Right: SAF3R outperforms prior efficient F3R meth￾ods on camera pose estimation and achieves performance close to the original model. transformer architecture [29] that jointly performs multiple geometric tasks, in￾cluding camera pose estimation and dense point map prediction. A key character￾istic … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the architecture shared by existing F3R transformers [12,15,31,36]. Image patches are first encoded by a DINO [5, 17] encoder and then processed by N alternating attention blocks of local and global attention, each consisting of multi-head self-attention (MHA) and feed-forward networks (FFN). lored sparse attention kernels for each. Building on this design, SAF3R adopts a two-stage pipeline con… view at source ↗
Figure 3
Figure 3. Figure 3: Representative global attention patterns across layers (G) and heads (H). Gray dashed lines in the attention maps separate tokens from different images. Similar ob￾servations apply to MapAnything and π 3 (see the appendix). Zoom in for best view. grouping to reduce similarity-computation overhead. However, merging-based methods heavily depend on anchor token selection, typically operate at relatively low s… view at source ↗
Figure 4
Figure 4. Figure 4: (a) and (b) illustrate the high sparsity of correspondence heads, while (c) shows the content-dependent dynamic nature of vertical-line heads. with distinct functional roles and sparsity patterns. In shallow layers, attention mainly performs local aggregation or focuses on a few salient patches, showing concentrated and often static sparsity with limited cross-view interaction. In middle layers, attention … view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the proposed SAF3R framework [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Efficiency results for processing a single scene with different sequence lengths. Latency and memory are measured on scenes with fixed image resolution 480 × 640. 0 2 4 6 8 10 12 14 16 18 20 22 Global Block Index 0 2 4 6 8 10 12 14 Head Index VGGT Profile 0 2 4 6 8 10 12 14 16 Global Block Index 0 2 4 6 8 10 12 14 Head Index 3 Profile 0 2 4 6 8 10 12 Global Block Index 0 3 6 9 12 15 18 21 Head Index DA3 Pr… view at source ↗
Figure 7
Figure 7. Figure 7: Offline head profiling results for different F3R models. After profiling, each global attention head is assigned to one of four predefined kernel types. sparse attention, highlighting the need for head-wise heterogeneous kernel con￾figurations and the flexibility of our method. 5.3 Ablation Study Effectiveness of tailored sparse kernels and offline search. We evaluate the effectiveness of the proposed offl… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison across different values of K for the DINO Top-K kernel on the Oxford Spires dataset. DA3 on a challenging large-scale outdoor scene with loop closures from the Ox￾ford Spires dataset [27]. As shown in [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative global attention patterns across layers (G) and heads (H) in four F3R transformers on the 7Scenes Chess scene. Gray dashed lines in the attention maps separate tokens from different images. Zoom in for best view. Full visualization results of four F3R transformers [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative global attention patterns across layers (G) and heads (H) in four F3R transformers on the 7Scenes Fire scene. Gray dashed lines in the attention maps separate tokens from different images. Zoom in for best view. Anything 3 (DA3) [15]. As shown, these models exhibit similar three-stage at￾tention patterns, consistent with the observations discussed in the main paper. Nevertheless, nuanced di… view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of the Top-10 selected keys from a correspondence head (VGGT￾G14-H4) and those selected based on DINO patch similarity. As shown, DINO patch similarity exhibits a high overlap with the patches selected by correspondence heads, especially covering the corresponding patches at the same spatial locations, motivating the use for sparse attention in these heads. B More Experimental Results Evaluatio… view at source ↗
Figure 12
Figure 12. Figure 12: Camera pose estimation results on Co3D-v2 under different sparsity settings. Flexibility of the proposed offline profiling method. Since the search pro￾cedure in the proposed offline profiling stage can start from any baseline sparsity level and progressively replace kernels to increase sparsity while maintaining ac￾curacy, we can initialize the process with different sparsity baselines to obtain models w… view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparisons of camera pose estimation and point cloud recon￾struction for different efficient methods on VGGT. As shown, the proposed SAF3R method produces accurate camera poses and high-quality point clouds while achieving faster inference than other methods [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparisons of camera pose estimation and point cloud recon￾struction between the original F3R transformers and our method. As shown, the pro￾posed SAF3R demonstrates strong generality across different F3R models and effec￾tively preserves the quality of the reconstructed point clouds [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] 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

49 extracted references · 17 linked inside Pith

  1. [1]

    IJCV120(2), 153–168 (2016)

    Aanæs, H., Jensen, R.R., Vogiatzis, G., Tola, E., Dahl, A.B.: Large-scale data for multiple-view stereopsis. IJCV120(2), 153–168 (2016)

    2016

  2. [2]

    In: Proceedings of the 29th ACM international conference on architectural support for programming languages and operating systems, volume 2

    Ansel,J.,Yang,E.,He,H.,Gimelshein,N.,Jain,A.,Voznesensky,M.,Bao,B.,Bell, P., Berard, D., Burovski, E., et al.: Pytorch 2: Faster machine learning through dy- namic python bytecode transformation and graph compilation. In: Proceedings of the 29th ACM international conference on architectural support for programming languages and operating systems, volume 2...

    2024

  3. [3]

    arXiv preprint arXiv:2210.09461 (2022)

    Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461 (2022)

    Pith/arXiv arXiv 2022

  4. [4]

    In: CVPRW

    Bolya, D., Hoffman, J.: Token merging for fast stable diffusion. In: CVPRW. pp. 4599–4603 (2023)

    2023

  5. [5]

    In: ICCV

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: ICCV. pp. 9650– 9660 (2021)

    2021

  6. [6]

    arXiv preprint arXiv:2511.14751 (2025)

    Chen, Y., Qiu, Y., Li, R., Agha, A., Omidshafiei, S., Patrikar, J., Scherer, S.: Co-me: Confidence-guided token merging for visual geometric transformers. arXiv preprint arXiv:2511.14751 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    In: CVPR

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828–5839 (2017)

    2017

  8. [8]

    arXiv preprint arXiv:2309.16588 (2023)

    Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need reg- isters. arXiv preprint arXiv:2309.16588 (2023)

    Pith/arXiv arXiv 2023

  9. [9]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  10. [10]

    Foundations and Trends in Computer Graphics and Vision9(1-2), 1–148 (2015)

    Furukawa, Y., Hernández, C.: Multi-view stereo: A tutorial. Foundations and Trends in Computer Graphics and Vision9(1-2), 1–148 (2015)

    2015

  11. [11]

    NeurIPS37, 52481–52515 (2024)

    Jiang, H., Li, Y., Zhang, C., Wu, Q., Luo, X., Ahn, S., Han, Z., Abdi, A.H., Li, D., Lin, C.Y., et al.: Minference 1.0: Accelerating pre-filling for long-context llms via dynamic sparse attention. NeurIPS37, 52481–52515 (2024)

    2024

  12. [12]

    arXiv preprint arXiv:2509.13414 (2025)

    Keetha, N., Müller, N., Schönberger, J., Porzi, L., Zhang, Y., Fischer, T., Knapitsch, A., Zauss, D., Weber, E., Antunes, N., et al.: Mapanything: Univer- sal feed-forward metric 3d reconstruction. arXiv preprint arXiv:2509.13414 (2025)

    Pith/arXiv arXiv 2025

  13. [13]

    arXiv preprint arXiv:2502.20766 (2025)

    Lai, X., Lu, J., Luo, Y., Ma, Y., Zhou, X.: Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference. arXiv preprint arXiv:2502.20766 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    In: ECCV

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

    2024

  15. [15]

    arXiv preprint arXiv:2511.10647 (2025)

    Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025)

    Pith/arXiv arXiv 2025

  16. [16]

    IEEE Transactions on Robotics33(5), 1255–1262 (2017)

    Mur-Artal, R., Tardós, J.D.: Orb-slam2: An open-source slam system for monoc- ular, stereo, and rgb-d cameras. IEEE Transactions on Robotics33(5), 1255–1262 (2017)

    2017

  17. [17]

    arXiv preprint arXiv:2304.07193 (2023) SAF3R 17

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023) SAF3R 17

    Pith/arXiv arXiv 2023

  18. [18]

    In: ICCV

    Reizenstein,J.,Shapovalov,R.,Henzler,P.,Sbordone,L.,Labatut,P.,Novotny,D.: Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In: ICCV. pp. 10901–10911 (2021)

    2021

  19. [19]

    arXiv preprint arXiv:2603.08055 (2026)

    Ren, W., Tan, X., Han, K.: Speed3r: Sparse feed-forward 3d reconstruction models. arXiv preprint arXiv:2603.08055 (2026)

    arXiv 2026

  20. [20]

    In: CVPR

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR. pp. 4104–4113 (2016)

    2016

  21. [21]

    In: ECCV

    Schönberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: ECCV. pp. 501–518. Springer (2016)

    2016

  22. [22]

    In: CVPR

    Schops, T., Schonberger, J.L., Galliani, S., Sattler, T., Schindler, K., Pollefeys, M., Geiger, A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In: CVPR. pp. 3260–3269 (2017)

    2017

  23. [23]

    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

  24. [24]

    In: CVPR

    Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in rgb-d images. In: CVPR. pp. 2930–2937 (2013)

    2013

  25. [25]

    arXiv preprint arXiv:2512.04939 (2025)

    Shu, Z., Lin, C., Xie, T., Yin, W., Li, B., Pu, Z., Li, W., Yao, Y., Cao, X., Guo, X., et al.: Litevggt: Boosting vanilla vggt via geometry-aware cached token merging. arXiv preprint arXiv:2512.04939 (2025)

    arXiv 2025

  26. [26]

    arXiv preprint arXiv:2512.02541 (2025)

    Sun,X.,Zhu,Z.,Lou,Z.,Yang,B.,Tang,J.,Zhang,L.,Wang,H.,Zhang,J.:Avggt: Rethinking global attention for accelerating vggt. arXiv preprint arXiv:2512.02541 (2025)

    arXiv 2025

  27. [27]

    The International Journal of Robotics Research 45(6), 839–857 (2026)

    Tao, Y., Muñoz-Bañón, M.Á., Zhang, L., Wang, J., Fu, L.F.T., Fallon, M.: The ox- ford spires dataset: Benchmarking large-scale lidar-visual localisation, reconstruc- tion and radiance field methods. The International Journal of Robotics Research 45(6), 839–857 (2026)

    2026

  28. [28]

    In: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages

    Tillet, P., Kung, H.T., Cox, D.: Triton: an intermediate language and compiler for tiled neural network computations. In: Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages. pp. 10–19 (2019)

    2019

  29. [29]

    NeurIPS30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. NeurIPS30(2017)

    2017

  30. [30]

    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

  31. [31]

    In: CVPR

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: CVPR. pp. 5294–5306 (2025)

    2025

  32. [32]

    In: CVPR

    Wang, J., Karaev, N., Rupprecht, C., Novotny, D.: Vggsfm: Visual geometry grounded deep structure from motion. In: CVPR. pp. 21686–21697 (2024)

    2024

  33. [33]

    In: CVPR

    Wang, Q., Zhang, Y., Holynski, A., Efros, A.A., Kanazawa, A.: Continuous 3d perception model with persistent state. In: CVPR. pp. 10510–10522 (2025)

    2025

  34. [34]

    In: CVPR

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: CVPR. pp. 20697–20709 (2024)

    2024

  35. [35]

    arXiv preprint arXiv:2511.21317 (2025)

    Wang, W., Meiner, L., Shubham, R., De La Parra, C., Kumar, A.: Httm: Head-wise temporal token merging for faster vggt. arXiv preprint arXiv:2511.21317 (2025)

    arXiv 2025

  36. [36]

    arXiv preprint arXiv:2507.13347 (2025)

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π3: Permutation-Equivariant Visual Geometry Learning. arXiv preprint arXiv:2507.13347 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    arXiv preprint arXiv:2512.01540 (2025) 18 J

    Wang, Z., Xu, D.: Flashvggt: Efficient and scalable visual geometry transformers with compressed descriptor attention. arXiv preprint arXiv:2512.01540 (2025) 18 J. Deng et al

    arXiv 2025

  38. [38]

    arXiv preprint arXiv:2502.01776 (2025)

    Xi, H., Yang, S., Zhao, Y., Xu, C., Li, M., Li, X., Lin, Y., Cai, H., Zhang, J., Li, D., et al.: Sparse videogen: Accelerating video diffusion transformers with spatial- temporal sparsity. arXiv preprint arXiv:2502.01776 (2025)

    Pith/arXiv arXiv 2025

  39. [39]

    arXiv preprint arXiv:2410.10819 (2024)

    Xiao, G., Tang, J., Zuo, J., Guo, J., Yang, S., Tang, H., Fu, Y., Han, S.: Duoatten- tion: Efficient long-context llm inference with retrieval and streaming heads. arXiv preprint arXiv:2410.10819 (2024)

    Pith/arXiv arXiv 2024

  40. [40]

    arXiv preprint arXiv:2309.17453 (2023)

    Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M.: Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453 (2023)

    Pith/arXiv arXiv 2023

  41. [41]

    In: ICML (2025)

    Xu, R., Xiao, G., Huang, H., Guo, J., Han, S.: Xattention: Block sparse attention with antidiagonal scoring. In: ICML (2025)

    2025

  42. [42]

    In: CVPR

    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: CVPR. pp. 21924–21935 (2025)

    2025

  43. [43]

    In: CVPR

    Yang,L.,Kang,B.,Huang,Z.,Xu,X.,Feng,J.,Zhao,H.:Depthanything:Unleash- ing the power of large-scale unlabeled data. In: CVPR. pp. 10371–10381 (2024)

    2024

  44. [44]

    arXiv preprint arXiv:2505.18875 (2025)

    Yang, S., Xi, H., Zhao, Y., Li, M., Zhang, J., Cai, H., Lin, Y., Li, X., Xu, C., Peng, K., et al.: Sparse videogen2: Accelerate video generation with sparse attention via semantic-aware permutation. arXiv preprint arXiv:2505.18875 (2025)

    Pith/arXiv arXiv 2025

  45. [45]

    In: ICCV

    Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: ICCV. pp. 12–22 (2023)

    2023

  46. [46]

    In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Yuan, J., Gao, H., Dai, D., Luo, J., Zhao, L., Zhang, Z., Xie, Z., Wei, Y., Wang, L., Xiao, Z., et al.: Native sparse attention: Hardware-aligned and natively trainable sparse attention. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 23078–23097 (2025)

    2025

  47. [47]

    In: ICML (2025)

    Zhang, J., Xiang, C., Huang, H., Wei, J., Xi, H., Zhu, J., Chen, J.: Spargeattn: Accurate sparse attention accelerating any model inference. In: ICML (2025)

    2025

  48. [48]

    arXiv preprint arXiv:2502.04507 (2025)

    Zhang, P., Chen, Y., Su, R., Ding, H., Stoica, I., Liu, Z., Zhang, H.: Fast video generation with sliding tile attention. arXiv preprint arXiv:2502.04507 (2025)

    Pith/arXiv arXiv 2025

  49. [49]

    9:Representative global attention patterns across layers (G) and heads (H) in four F3R transformers on the7Scenes Chessscene

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ingviewsynthesisusingmultiplaneimages.arXivpreprintarXiv:1805.09817(2018) SAF3R 19 Appendix A Additional Results on Attention Analysis G14 - H10 G14 - H4 G14 - H13 G0 - H3 G0 - H7 Global Attention Maps ( VGGT ) G0 - H1 Global Attention Maps ( Depth Anything 3 ) Shallow D...

    Pith/arXiv arXiv 2018