Pith. sign in

REVIEW 3 major objections 6 minor 51 references

Anchoring 3D Gaussians to ray maps stops pose drift in long sequences

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 · glm-5.2

2026-07-09 18:22 UTC pith:JDSZEO4Q

load-bearing objection Real method, confounded headline: RGC works but backbone scaling may deserve more credit than the paper gives it. the 3 major comments →

arxiv 2607.07168 v1 pith:JDSZEO4Q submitted 2026-07-08 cs.CV

NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction

Xiangyu Sun , Liu Liu , Seungkwon Yang , Jingbing Han , Seungtae Nam , Zhizhong Su , Eunbyung Park This is my paper
classification cs.CV
keywords geometryposereconstructionfeed-forwardpose-freerenderingacrossappearance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the dominant failure mode in pose-free feed-forward 3D Gaussian Splatting is cumulative camera-pose drift over long image sequences, and that this drift arises because geometry and appearance are optimized in isolation. The authors propose to tie the two together by deriving each Gaussian's 3D center directly from a per-pixel ray map (direction plus origin) and a depth value, so that rendering losses back-propagate into geometric predictions and geometric losses constrain the rendered output. This bidirectional coupling, trained under a dual-frequency curriculum that interleaves easy high-overlap pairs with progressively harder wide-baseline ones, is shown to suppress drift and improve both novel-view synthesis and camera-pose accuracy, with gains widening as sequence length increases.

Core claim

The central mechanism is a raymap-guided lifting equation: each pixel's 3D Gaussian center is computed as the ray origin plus depth times the ray direction. Because the Gaussian position is an explicit function of the predicted ray map and depth, an RGB rendering loss gradients flow back into the ray-map and depth predictions, while a separate ray-map consistency loss constrains the geometry that determines rendering. The paper's ablations show that removing the ray-map loss degrades rendering quality and removing the RGB loss degrades pose accuracy, supporting the claim that each supervision signal refines the other. The dual-frequency scheduling component addresses a training instability:纯

What carries the argument

Three components carry the argument. First, the Raymap-Guided Coupling Module lifts pixels to 3D using a predicted per-pixel ray origin and direction combined with a depth scalar, making Gaussian positions a direct function of geometric predictions rather than independent parameters. Second, a unified loss jointly optimizes RGB reconstruction, ray-map consistency, and camera-parameter regularization, so that gradients from each term flow through the shared ray-map representation. Third, a Dual-Frequency Viewpoint Scheduling strategy pairs an easy-to-hard overlap curriculum with stochastic replay of small-interval (high-overlap) samples, counteracting a tendency for curriculum expansion to er

Load-bearing premise

The framework still relies on SfM-based pseudo-ground-truth camera poses and ray maps as supervision targets during training, even though the authors acknowledge that SfM introduces sensor noise. If those pseudo-labels are systematically biased in particular scene types, the model may learn to reproduce that bias rather than achieve true geometric consistency.

What would settle it

If one could decouple the bidirectional gradient flow—for instance, by stopping gradients from the RGB loss into the ray-map predictions while keeping all other components identical—and the model still showed equivalent drift suppression and rendering gains, then the central claim that explicit geometry-appearance coupling is the causal mechanism would be undermined.

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

If this is right

  • If the bidirectional coupling claim holds generally, explicit geometric anchoring of scene representations could replace loosely connected pose-and-appearance pipelines in other feed-forward 3D reconstruction settings beyond Gaussian Splatting.
  • The finding that replay of easy samples is necessary to prevent short-range degradation during curriculum expansion suggests that current pose-free training schedules trade local geometric consistency for long-range coverage, a trade-off that may recur in other multi-view learning tasks.
  • The widening performance gap at longer sequences (12v, 24v) implies that drift accumulation, not representation capacity, is the binding constraint in pose-free feed-forward 3DGS, which would redirect architectural effort toward geometric coupling rather than model scaling.
  • Cross-dataset pose transfer results (trained on outdoor scenes, tested on indoor) suggest that ray-map supervision produces pose representations that generalize beyond the training domain, which could reduce the need for domain-specific pose annotation.

Where Pith is reading between the lines

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

  • If ray-map consistency is the conduit through which appearance supervision refines pose, then the quality of the ray-map pseudo-ground-truth during training sets an upper bound on achievable pose accuracy; systematic biases in the training-time geometry teacher would propagate into the student model.
  • The dual-frequency scheduling result (replay prevents short-range collapse while curriculum expansion improves long-range coverage) may reflect a general plasticity-stability trade-off in multi-view curricula that could be tested in other settings such as video-based novel-view synthesis or multi-view stereo training.
  • The paper's reliance on SfM-derived poses and ray maps during training, despite noting their sensor noise, leaves open whether a self-supervised variant that dispenses with pseudo-ground-truth geometry entirely could retain the same coupling benefits.

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 / 6 minor

Summary. The paper proposes NoDrift3R, a pose-free feed-forward 3D Gaussian Splatting framework that addresses cumulative pose drift in long image sequences. The central technical contribution is a Raymap-Guided Coupling (RGC) module (§3.1, Eq. 1) that anchors Gaussian centers to raymap-derived geometry (ray direction and origin combined with depth), creating a bidirectional optimization loop between appearance supervision and geometric constraints. A Dual-Frequency Viewpoint Scheduling strategy (§3.3) combines easy-to-hard interval expansion with stochastic replay of small-interval pairs. Experiments on DL3DV, RE10K, and ScanNet++ show improvements over prior pose-free methods (YoNoSplat, NoPoSplat, AnySplat) in both novel-view synthesis and pose estimation, with larger gains at longer sequences (24v).

Significance. The paper addresses a practically important problem—pose drift accumulation in long-sequence pose-free 3DGS—and proposes a conceptually clean solution. The RGC mechanism (Eq. 1) is a parameter-free derivation of Gaussian positions from predicted raymaps and depth, and the bidirectional coupling is empirically validated by ablation (Table 6): removing raymap loss degrades rendering, removing RGB loss degrades pose accuracy. The Dual-Frequency scheduling ablation (Table 7) provides actionable insight about the short-range/long-range trade-off. Cross-dataset generalization results (Tables 3, 5) and zero-shot transfer are commendable. The framework is built on established backbones (Depth Anything v3, DINOv2), which aids reproducibility but also raises attribution questions (see Major Comments).

major comments (3)
  1. The headline SOTA comparisons in Tables 1–2 use the Giant model (Depth Anything v3, hidden dim 1536, 40 blocks) against baselines (YoNoSplat, NoPoSplat, AnySplat) that employ different and potentially weaker backbones. The paper's own ablation data in Table 6 (Large model) shows that RGC contributes +0.053 AUC@5° at 6v (0.821→0.874) and +0.097 at 24v (0.690→0.787), while scaling from Large to Giant (both with RGC) adds +0.093 AUC@5° at 6v (0.874→0.967) and +0.162 at 24v (0.787→0.949). Backbone scaling thus contributes 1.5–1.7× more than RGC to the final pose numbers. Without equipping at least one baseline with the same Depth Anything v3 backbone, the reported gains over SOTA cannot be cleanly attributed to RGC versus backbone capacity. The paper's claim that RGC is 'the key' (Abstract, §3.2) is partially undermined by its own ablation data. The authors should either (a) re-run at least
  2. Table 6 reports single-run ablation results without error bars or statistical significance tests. The RGC contribution at 6v is +0.24 PSNR (23.062→23.302) and +0.053 AUC@5° (0.821→0.874). Given the moderate magnitude of these deltas, particularly for PSNR, it is unclear whether they are stable across random seeds. The central claim of bidirectional coupling rests on these ablation results; reporting variance over at least 3 runs would substantially strengthen the conclusion.
  3. §3.2, Eqs. (3)–(6): The loss weights λ_mse, λ_lpips, λ_cam, λ_ray are listed as free parameters in the axiom ledger but their values are not reported in the main text. Since the total objective (Eq. 6) is a weighted sum and the relative weighting between RGB, camera, and raymap losses directly governs the 'bidirectional feedback loop' that is the paper's central mechanism, these values are load-bearing for reproducibility and should be reported.
minor comments (6)
  1. §3.1, Eq. (1): The notation R ∈ R^{HW×6} is unusual; clarifying that it is reshaped as [r | o] where r, o ∈ R^{HW×3} per-pixel would help readers.
  2. §3.3: The overlap threshold schedule anneals from 1.0 to 0.75, but the functional form (linear, cosine, stepwise, etc.) is not specified.
  3. Table 1: The 'Erayzer256×256' entry reports 24.814 PSNR at 6v but drops sharply to 18.750 at 24v, while the text does not discuss this degradation pattern. A brief note would contextualize the comparison.
  4. Figure 2 is referenced as illustrating the synergistic framework but the sub-figures (a), (b), (c) are discussed in the text without clear visual labels in the caption.
  5. §4.1: The paper states the Giant model is trained on 8 H100 GPUs and the Large model on 8 RTX5090 GPUs; the RTX5090 may be a typo and should be verified.
  6. Table 7: The 'original sampler' baseline is described as 'similar to YoNoSplat' but the exact configuration is not specified, making reproduction difficult.

Circularity Check

0 steps flagged

No circularity found: the derivation chain is self-contained and independently grounded

full rationale

The paper's core derivation chain is not circular. Eq. 1 (p_j = o_j + D_j · r_j) is a standard, parameter-free ray unprojection operation — it does not define its output in terms of the quantity it claims to produce. The three losses (Eqs. 3–5: RGB MSE+LPIPS, camera Huber, raymap L1) are standard supervised losses with independently defined targets (rendered images, SfM pseudo-GT poses, pseudo-GT raymaps). The claimed 'bidirectional feedback loop' is an architectural property of the computation graph (Gaussian positions depend on raymaps, so RGB loss gradients flow into raymap predictions, and raymap loss constrains Gaussian positions), not a derived result that reduces to its inputs by definition. The ablation in Table 6 independently tests each loss component's contribution by removal, showing measurable degradation — this is empirical validation, not circular reasoning. Self-citations exist (Uni3R [33] shares the first author; iLRM [16] and MVP [17] share co-author Park), but these are used as related work or baselines, not as load-bearing premises for the central claim. The backbone (Depth Anything v3 [23]) has no author overlap with this paper. The skeptic's concern about confounding RGC with backbone capacity is a valid experimental attribution issue, but it is not circularity — the paper does not define RGC's contribution in terms of its own evaluation results. No step in the derivation chain reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 2 axioms · 0 invented entities

The paper introduces standard loss weights and scheduling hyperparameters but no new theoretical entities or postulated physical constructs.

free parameters (7)
  • lambda_mse
    Weight for MSE loss in Eq. 3, value not specified in text.
  • lambda_lpips
    Weight for LPIPS loss in Eq. 3, value not specified.
  • lambda_cam
    Weight for camera loss in Eq. 6, value not specified.
  • lambda_ray
    Weight for raymap loss in Eq. 6, value not specified.
  • g_max = ,
    Maximum view interval clip, set to 15 in Sec. 3.3.
  • p_small = ,
    Probability of sampling small-interval replay, set to 0.5 in Sec. 3.3.
  • overlap_threshold_schedule = 1.0 to 0.75
    Annealing schedule for overlap target threshold in Sec. 3.3.
axioms (2)
  • domain assumption SfM pseudo-ground-truth poses are sufficiently accurate to supervise camera and raymap predictions.
    The paper notes SfM noise but still uses it as ground truth for L_cam and L_ray (Eqs. 4, 5).
  • domain assumption DINOv2 cosine similarity is a valid proxy for visual overlap.
    Used in Eq. 7 to schedule training pairs based on overlap.

pith-pipeline@v1.1.0-glm · 18055 in / 1968 out tokens · 295740 ms · 2026-07-09T18:22:11.862236+00:00 · methodology

0 comments
read the original abstract

Pose-Free Feed-forward 3D Gaussian Splatting (3DGS) has recently emerged as a powerful paradigm for fast scene reconstruction. However, its performance degrades significantly in long image sequences due to cumulative camera pose estimation drift, which propagates errors into geometric modeling and severely limits rendering fidelity. In this work, we revisit the long-sequence bottleneck and identify pose drift as the primary factor restricting reconstruction quality. Furthermore, while SfM-based pseudo ground-truth poses introduce sensor noise, purely rendering-based supervision often leads to optimization instability and local minima due to the entangled optimization of geometry and pose. To address the challenges, we propose a synergistic pose-free framework that explicitly couples geometry and appearance via a Raymap-Guided Coupling Module (RGC). Concretely, we anchor Gaussian centers to raymap-induced geometry and jointly optimize RGB reconstruction, raymap consistency, and camera regularization under a unified objective, yielding a bidirectional feedback loop: stronger geometry improves rendering, and appearance supervision in turn refines geometry and pose. To further stabilize learning across wide temporal ranges, we introduce a Dual-Frequency Viewpoint Scheduling strategy that combines easy-to-hard interval expansion with replay of short-interval pairs. Extensive experiments across in-domain and cross-domain datasets show consistent gains in both rendering and pose estimation, with notably improved robustness on long sequences. Ablation studies validate our central insight: explicitly designed geometry-appearance synergy is the key to scalable and drift-robust pose-free feed-forward 3D reconstruction.

Figures

Figures reproduced from arXiv: 2607.07168 by Eunbyung Park, Jingbing Han, Liu Liu, Seungkwon Yang, Seungtae Nam, Xiangyu Sun, Zhizhong Su.

Figure 1
Figure 1. Figure 1: Overview of our synergistic pose-free framework for feed-forward 3D recon￾struction. Our method effectively suppresses the pose drift problem, especially in long￾sequence settings. Left: representative failure cases of existing methods under pose drift. Right: our pipeline and outputs (camera poses and 3D Gaussians). scene representations directly from multi-view images. Compared to conven￾tional per-scene… view at source ↗
Figure 2
Figure 2. Figure 2: Our synergistic framework for "Rendering-to-Geometry Gain". – A Synergistic Framework for "Rendering-to-Geometry Gain": We propose a novel pose-free framework that integrates 3DGS rendering supervision with explicit raymap constraints. Unlike previous methods that optimize geometry and pose in isolation, our approach establishes a positive feedback loop—which we term "Rendering-to-Geometry Gain"—where appe… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Synergistic Framework. (a) Ray-Based Feed-Forward GS Model. We employ a single transformer (Vallina DINOv2 model), followed by Gaussian Head, Depth Head, Ray-Map Head, and Camera Head. (b) A Raymap-Guided Coupling Module for "Rendering-to-Geometry Gain". (c) Our Replay & Overlap Scheduler leads to robust performance across arbitrary intervals. 3 Method In this section, we first propose our … view at source ↗
Figure 4
Figure 4. Figure 4: Pose visualization and compared with representative methods. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of novel view synthesis on the DL3DV test [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: As the number of input views increases (6→12→24), YoNoSplat degrades markedly, while our method remains stable across all metrics, demonstrating stronger robustness to pose drift in long sequences [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Our model generalizes better than YoNoSplat on the ScanNet++ dataset and demonstrates coherent fusion of Gaussians across longer sequences. We attribute this gain to our synergistic framework, which tightly couples geometry and appearance. Cross-Dataset Generalization To evaluate cross-dataset generalization, we train the model on DL3DV and directly evaluate on ScanNet++ without any fine-tuning. We compare… 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

51 extracted references · 51 canonical work pages · 18 internal anchors

  1. [1]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Charatan,D.,Li,S.L.,Tagliasacchi,A.,Sitzmann,V.:pixelsplat:3dgaussiansplats from image pairs for scalable generalizable 3d reconstruction. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 19457–19467 (2024) 2, 4

    work page 2024

  2. [2]

    MVSNeRF: Fast Generalizable Radiance Field Reconstruction from Multi-View Stereo

    Chen, A., Xu, Z., Zhao, F., Zhang, X., Xiang, F., Yu, J., Su, H.: Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. arXiv preprint arXiv:2103.15595 (2021) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  3. [3]

    In: European conference on computer vision

    Chen, Y., Xu, H., Zheng, C., Zhuang, B., Pollefeys, M., Geiger, A., Cham, T.J., Cai, J.: Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In: European conference on computer vision. pp. 370–386. Springer (2024) 2, 4, 10, 12

    work page 2024

  4. [4]

    Advances in Neural Information Processing Systems37, 107064–107086 (2024) 4

    Chen, Y., Zheng, C., Xu, H., Zhuang, B., Vedaldi, A., Cham, T.J., Cai, J.: Mvs- plat360: Feed-forward 360 scene synthesis from sparse views. Advances in Neural Information Processing Systems37, 107064–107086 (2024) 4

    work page 2024

  5. [5]

    Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality

    Dao, T., Gu, A.: Transformers are ssms: Generalized models and efficient algo- rithms through structured state space duality. arXiv preprint arXiv:2405.21060 (2024) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    arXiv preprint arXiv:2512.08930 (2025) 5

    Deng, Y., Peng, S., Zhang, J., Heal, K., Sun, T., Flynn, J., Marschner, S., Chai, L.: Selfi: Self improving reconstruction engine via 3d geometric feature alignment. arXiv preprint arXiv:2512.08930 (2025) 5

    work page arXiv 2025

  7. [7]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    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) 4, 9

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  8. [8]

    InstantSplat: Sparse-view Gaussian Splatting in Seconds

    Fan, Z., Cong, W., Wen, K., Wang, K., Zhang, J., Ding, X., Xu, D., Ivanovic, B., Pavone, M., Pavlakos, G., et al.: Instantsplat: Sparse-view gaussian splatting in seconds. arXiv preprint arXiv:2403.20309 (2024) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition

    Fu, Y., Liu, S., Kulkarni, A., Kautz, J., Efros, A.A., Wang, X.: Colmap-free 3d gaussian splatting. In: Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition. pp. 20796–20805 (2024) 5

    work page 2024

  10. [10]

    PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting

    Hong, S., Jung, J., Shin, H., Han, J., Yang, J., Luo, C., Kim, S.: Pf3plat: Pose-free feed-forward 3d gaussian splatting. arXiv preprint arXiv:2410.22128 (2024) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    LRM: Large Reconstruction Model for Single Image to 3D

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  12. [12]

    In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision

    Huang, R., Mikolajczyk, K.: No pose at all: Self-supervised pose-free 3d gaussian splatting from sparse views. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision. pp. 27947–27957 (2025) 5

    work page 2025

  13. [13]

    In: Proceedings of the SIG- GRAPH Asia 2025 Conference Papers

    Imtiaz, T., Chai, L., Heal, K., Luo, X., Park, J., Dy, J., Flynn, J.: Lvt: Large- scale scene reconstruction via local view transformers. In: Proceedings of the SIG- GRAPH Asia 2025 Conference Papers. pp. 1–12 (2025) 5

    work page 2025

  14. [14]

    ACM Transactions on Graphics (TOG)44(6), 1–16 (2025) 2, 3, 5, 7, 10, 12 NoDrift3R 17

    Jiang, L., Mao, Y., Xu, L., Lu, T., Ren, K., Jin, Y., Xu, X., Yu, M., Pang, J., Zhao, F., et al.: Anysplat: Feed-forward 3d gaussian splatting from unconstrained views. ACM Transactions on Graphics (TOG)44(6), 1–16 (2025) 2, 3, 5, 7, 10, 12 NoDrift3R 17

    work page 2025

  15. [15]

    LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias

    Jin, H., Jiang, H., Tan, H., Zhang, K., Bi, S., Zhang, T., Luan, F., Snavely, N., Xu, Z.: Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242 (2024) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  16. [16]

    iLRM: An Iterative Large 3D Reconstruction Model

    Kang, G., Nam, S., Yang, S., Sun, X., Khamis, S., Mohamed, A., Park, E.: ilrm: An iterative large 3d reconstruction model. arXiv preprint arXiv:2507.23277 (2025) 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Multi-view Pyramid Transformer: Look Coarser to See Broader

    Kang, G., Yang, S., Nam, S., Lee, Y., Kim, J., Park, E.: Multi-view pyramid transformer: Look coarser to see broader. arXiv preprint arXiv:2512.07806 (2025) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Kang, G., Yoo, J., Park, J., Nam, S., Im, H., Shin, S., Kim, S., Park, E.: Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splatting. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 22012–22022 (2025) 2, 5

    work page 2025

  19. [19]

    MapAnything: Universal Feed-Forward Metric 3D Reconstruction

    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) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  20. [20]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023) 2, 4

    work page 2023

  21. [21]

    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) 5

    work page 2024

  22. [22]

    Lin, C.H., Ma, W.C., Torralba, A., Lucey, S.: Barf: Bundle-adjusting neural radi- ancefields.In:ProceedingsoftheIEEE/CVFinternationalconferenceoncomputer vision. pp. 5741–5751 (2021) 5

    work page 2021

  23. [23]

    Depth Anything 3: Recovering the Visual Space from Any Views

    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) 5, 6, 9, 10

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Ling, L., Sheng, Y., Tu, Z., Zhao, W., Xin, C., Wan, K., Yu, L., Guo, Q., Yu, Z., Lu, Y., et al.: Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 22160–22169 (2024) 10

    work page 2024

  25. [25]

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

    Meng, Q., Chen, A., Luo, H., Wu, M., Su, H., Xu, L., He, X., Yu, J.: Gnerf: Gan-based neural radiance field without posed camera. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6351–6361 (2021) 5

    work page 2021

  26. [26]

    Commu- nications of the ACM65(1), 99–106 (2021) 2, 4

    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) 2, 4

    work page 2021

  27. [27]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Nam, S., Sun, X., Kang, G., Lee, Y., Oh, S., Park, E.: Generative densification: Learning to densify gaussians for high-fidelity generalizable 3d reconstruction. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 26683–26693 (2025) 2, 4

    work page 2025

  28. [28]

    DINOv2: Learning Robust Visual Features without Supervision

    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) 8, 9

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    Advances in Neural Information Processing Sys- tems32(2019) 9 18 X

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in Neural Information Processing Sys- tems32(2019) 9 18 X. Sun et al

    work page 2019

  30. [30]

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

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12179–12188 (2021) 6

    work page 2021

  31. [31]

    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) 3, 5, 10

    work page 2016

  32. [32]

    In: European conference on computer vision

    Schönberger, J.L., Zheng, E., Frahm, J.M., Pollefeys, M.: Pixelwise view selection for unstructured multi-view stereo. In: European conference on computer vision. pp. 501–518. Springer (2016) 10

    work page 2016

  33. [33]

    arXiv preprint arXiv:2508.03643 (2025) 2, 3, 5

    Sun, X., Jiang, H., Liu, L., Nam, S., Kang, G., Wang, X., Sui, W., Su, Z., Liu, W., Wang, X., et al.: Uni3r: Unified 3d reconstruction and semantic understanding via generalizable gaussian splatting from unposed multi-view images. arXiv preprint arXiv:2508.03643 (2025) 2, 3, 5

    work page arXiv 2025

  34. [34]

    Learning to (Learn at Test Time): RNNs with Expressive Hidden States

    Sun, Y., Li, X., Dalal, K., Xu, J., Vikram, A., Zhang, G., Dubois, Y., Chen, X., Wang, X., Koyejo, S., et al.: Learning to (learn at test time): Rnns with expressive hidden states. arXiv preprint arXiv:2407.04620 (2024) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    arXiv preprint arXiv:2602.20160 (2026) 5

    Wang, C., Tan, H., Yifan, W., Chen, Z., Liu, Y., Sunkavalli, K., Bi, S., Liu, L., Hu, Y.:tttlrm:Test-timetrainingforlongcontextandautoregressive3dreconstruction. arXiv preprint arXiv:2602.20160 (2026) 5

    work page arXiv 2026

  36. [36]

    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) 2, 5, 10, 12

    work page 2025

  37. [37]

    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) 5

    work page 2024

  38. [38]

    $\pi^3$: Permutation-Equivariant Visual Geometry Learning

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.: pi3: Permutation-equivariant visual geometry learning. arXiv preprint arXiv:2507.13347 (2025) 2, 5, 10, 12

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    IEEE transactions on image processing 13(4), 600–612 (2004) 9

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004) 9

    work page 2004

  40. [40]

    NeRF--: Neural Radiance Fields Without Known Camera Parameters

    Wang, Z., Wu, S., Xie, W., Chen, M., Prisacariu, V.A.: NeRF−−: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064 (2021) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  41. [41]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, H., Peng, S., Wang, F., Blum, H., Barath, D., Geiger, A., Pollefeys, M.: Depth- splat: Connecting gaussian splatting and depth. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16453–16463 (2025) 2, 4, 10, 12

    work page 2025

  42. [42]

    arXiv preprint arXiv:2511.07321 (2025) 2, 3, 5, 7, 10, 12, 14

    Ye, B., Chen, B., Xu, H., Barath, D., Pollefeys, M.: Yonosplat: You only need one model for feedforward 3d gaussian splatting. arXiv preprint arXiv:2511.07321 (2025) 2, 3, 5, 7, 10, 12, 14

    work page arXiv 2025

  43. [43]

    No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images

    Ye, B., Liu, S., Xu, H., Li, X., Pollefeys, M., Yang, M.H., Peng, S.: No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207 (2024) 2, 3, 5, 10, 12

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  44. [44]

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

    Yeshwanth, C., Liu, Y.C., Nießner, M., Dai, A.: Scannet++: A high-fidelity dataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12–22 (2023) 10

    work page 2023

  45. [45]

    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) 5 NoDrift3R 19

    work page 2024

  46. [46]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effec- tiveness of deep features as a perceptual metric. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 586–595 (2018) 9

    work page 2018

  47. [47]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhang, S., Wang, J., Xu, Y., Xue, N., Rupprecht, C., Zhou, X., Shen, Y., Wet- zstein, G.: Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 21936–21947 (2025) 5

    work page 2025

  48. [48]

    Test-Time Training Done Right

    Zhang, T., Bi, S., Hong, Y., Zhang, K., Luan, F., Yang, S., Sunkavalli, K., Freeman, W.T., Tan, H.: Test-time training done right. arXiv preprint arXiv:2505.23884 (2025) 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    arXiv preprint arXiv:2512.10950 (2025) 2, 5, 8, 12

    Zhao, Q., Tan, H., Wang, Q., Bi, S., Zhang, K., Sunkavalli, K., Tulsiani, S., Jiang, H.: E-rayzer: Self-supervised 3d reconstruction as spatial visual pre-training. arXiv preprint arXiv:2512.10950 (2025) 2, 5, 8, 12

    work page arXiv 2025

  50. [50]

    Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ingviewsynthesisusingmultiplaneimages.arXivpreprintarXiv:1805.09817(2018) 10

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  51. [51]

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

    Ziwen, C., Tan, H., Zhang, K., Bi, S., Luan, F., Hong, Y., Fuxin, L., Xu, Z.: Long- lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4349–4359 (2025) 5

    work page 2025