Pith. sign in

REVIEW 4 major objections 5 minor 65 references

A sparse 3D refiner fixes the local distortions that 2D decoders inject into single-image geometry.

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

2026-08-01 16:30 UTC pith:E7UFAUY3

load-bearing objection A well-ablated 3D-refinement architecture that plausibly improves fine-detail geometry, but the headline local evaluation protocol is too fragile to fully establish the central claim. the 4 major comments →

arxiv 2607.17967 v2 pith:E7UFAUY3 submitted 2026-07-20 cs.CV

MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement

Lingyu Kong , Ruicheng Li , Ruicheng Wang , Sicheng Xu , Chengtang Yao , Jianfeng Xiang , Jiaolong Yang This is my paper
classification cs.CV
keywords monocular geometry estimationpoint mapsparse 3D convolutionsdepth refinementzero-shot depththin structuresarchitectural mismatch3D voxel shell
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 current monocular geometry models distort fine 3D structure because they decode 3D geometry with 2D convolutions: features follow image-plane proximity, so pixels from different surfaces get mixed at depth edges. MoGe-3 keeps a strong 2D backbone but lifts its coarse point map onto a sparse 3D voxel shell, then iteratively refines log-depth with 3D sparse convolutions that only aggregate along true 3D neighborhoods. On zero-shot benchmarks the refinement improves both global accuracy and, most sharply, local fine-detail metrics such as δ0.01 on point maps (55.9 vs. 46.6 for MoGe-2). A parameter-matched 2D refiner does not produce these gains, which the paper takes as evidence for its architectural-mismatch diagnosis. If correct, the design reframes where the remaining fidelity bottleneck lies in monocular geometry estimation.

Core claim

MoGe-3 claims that fine-detail monocular geometry errors are largely an architectural artifact. Because standard decoders aggregate features by image-plane adjacency, geometrically distant surfaces that touch in the image contaminate each other, over-smoothing thin structures. SSR addresses this by voxelizing the predicted point map into a thin sparse shell using fixed image axes (u,v) and a quantized log-depth axis, then running a sparse 3D U-Net that aggregates features only among 3D-adjacent voxels, with the occupied voxel set re-derived from the latest prediction each iteration. The refiner predicts residual log-depth updates; after three iterations the default ViT-L model beats MoGe-2 o

What carries the argument

Self-Guided Sparse 3D Refiner (SSR): an iterative module that takes the current point map, voxelizes it as (i, j, quantized log-depth) to form a sparse voxel shell with exactly one voxel per pixel, processes it with a sparse 3D U-Net whose downsampling depends on scene geometry, injects 2D backbone features at the bottleneck, and adds the predicted log-depth residual to update the shell. Its two properties: projection-aligned, scale-invariant shell; geometry-coherent receptive field that decouples features across depth discontinuities.

Load-bearing premise

The headline fine-detail claim rests on local metrics computed with the authors' mask-selection and per-segment-alignment protocol on two synthetic datasets; if that protocol does not reflect real-world fine-detail 3D fidelity, the claim is not established for real images.

What would settle it

Evaluate the same models on a real-world high-fidelity dataset (e.g., ScanNet++-style or LiDAR scans of thin structures) using the paper's fine-detail masks but with only a single global-scale alignment — no per-segment translation — and compare local δ0.01 between MoGe-3 and MoGe-2. If the gap shrinks toward zero, the central claim fails; if it persists, the claim survives.

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

If this is right

  • If the diagnosis is right, adding 3D-aware refinement to any existing monocular depth or point-map model should improve local fidelity, because the bottleneck is the 2D decoding geometry, not the encoder.
  • The gains saturate and remain stable beyond the training iteration count (K=3 up to K=7), suggesting the residual updates are a convergent correction rather than overfitting to a fixed loop.
  • Voxel resolution has a clear optimum near D=200; too coarse misses fine structure, too fine starves the sparse convolutions of occupancy, guiding future implementations.
  • Global metrics also improve, so the refiner sharpens details without destabilizing the overall scene layout.
  • SSR remains effective on a stronger ViT-G backbone, so it is complementary to encoder scaling.

Where Pith is reading between the lines

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

  • The local evaluation protocol (per-segment translation with shared global scale) can absorb constant depth offsets within each segment, so the reported δ0.01 gains likely overstate real-world thin-structure accuracy at absolute depths; a variant that removes the per-segment shift would be a sharper test.
  • The paper's own boundary-F1 result (third, behind InfiniDepth) hints that image-space edge sharpness is not fully controlled by 3D refinement; combining SSR with pixel-space or generative decoding might push both 2D edges and 3D fidelity.
  • A natural extension is to apply SSR on real sensor data (e.g., LiDAR or high-res RGB-D) where fine-detail ground truth exists, to test whether the synthetic-data gains transfer.
  • The regression-based limitation acknowledged by the authors suggests the next step is hybrid 3D refinement with a generative boundary model to remove fly-points.

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 / 5 minor

Summary. MoGe-3 proposes a Self-Guided Sparse 3D Refinement (SSR) module attached to a MoGe-2 base model. The initial 2D point map is voxelized in a log-depth coordinate space, iteratively refined by a sparse 3D U-Net that aggregates features according to 3D spatial locality, and the predicted log-depth residuals update the point map. The paper claims this 3D inductive bias corrects an architectural mismatch in 2D decoders and significantly improves fine-detail geometry, reporting zero-shot global, local, and boundary metrics on nine datasets, together with ablations over voxel resolution, feature injection, 2D vs 3D refinement, and iteration count.

Significance. If the fine-detail claim holds, the paper is a meaningful step: it provides a clean architectural argument for 3D sparse refinement over 2D decoding and includes several well-controlled experiments. The parameter-matched 2D refiner ablation (Table 3), the iteration sweep beyond the training horizon (Fig. 4), and the honest D-sweep are strengths, and the promise of released code/models supports reproducibility. However, the headline local-accuracy result currently rests on a narrow evaluation protocol: two synthetic datasets, hand-set mask thresholds, and per-segment translation alignment. The paper needs additional validation or careful qualification before the abstract's claim of 'significantly outperforming existing approaches in recovering fine detailed 3D geometry' is fully established for real images.

major comments (4)
  1. [Appendix B.2, Eq. (18); Table 1/C.2] The central fine-detail claim is supported only by local metrics on Spring and Synth4K, computed after fitting a per-segment translation t_i (Eq. 18) for every SAM2 segment. This removes per-object depth offsets before computing Rel and δ0.01. For thin structures with limited internal depth variation, a smoothed or planar prediction can be translated to match the GT segment and still score well. Please report the same local metrics without per-segment translation, or with only a single global shift, and analyze the distribution of fitted shifts. As written, the local protocol may measure within-segment shape after discarding exactly the per-object depth errors that are part of 3D geometric fidelity.
  2. [Sec. 4.3; Table C.1 (metric depth block)] The text states that the method achieves 'best global performance across all three evaluation settings' and 'consistently outperforms' prior work. This is contradicted by per-dataset metric-depth results: on NYUv2 the ViT-L model gives Rel 8.43 vs MoGe-2's 6.90, and on ETH3D 14.8 vs 9.99, with corresponding δ1 regressions. The averaged metric-depth gains are driven by other datasets. This does not invalidate the method, but the wording overstates consistency. Report per-dataset variability and qualify the global claim.
  3. [Sec. 4.3; Table 1/C.3] Boundary F1 is the only real-data metric with a local component, and the method ranks third behind InfiniDepth and Depth Pro. The paper argues that F1 is a 2D pixel-level metric and does not capture 3D geometric quality. That argument is reasonable, but it also means the claim of superior fine-detail 3D recovery on real images has no direct quantitative support. Either add real-data local geometry evaluation (e.g., manually annotated or semi-automatic fine-structure masks on real dense depth) or explicitly limit the fine-detail claim to the synthetic benchmarks.
  4. [Sec. 3.3, A.3] The refiner receives gradients only from synthetic samples, while the base model sees the full real/synthetic mixture. This design choice is understandable, but it heightens the need for real-world fine-detail validation. The current evaluation does not demonstrate that the 3D refiner's synthetic-trained residual corrections transfer to real thin structures beyond what is visible in averaged global metrics and a few qualitative examples. Please provide additional real-data evidence or temper the real-image claim in the abstract and Sec. 5.
minor comments (5)
  1. [Sec. 4.2] For local metrics, the text says thresholds of 0.01 for point maps and 1.01 for depth maps, but Table 1 headings read simply 'δ0.01'. Clarify the depth threshold in the table or caption to avoid ambiguity.
  2. [Sec. 4.5, Table 3] The ablation models are trained with K=3 but evaluated at K=5, while the jointly trained model in Fig. 4 is trained with K=3 and evaluated up to K=7. State this explicitly in the ablation text; the current phrasing could be misread as training at K=5.
  3. [Sec. 3.2.1, Eq. (3)] The notation ⌊·⌉ is used for rounding but defined only in passing. Define it explicitly at first use.
  4. [Sec. 4.4] For Pixel-Perfect Depth and InfiniDepth, the paper uses MoGe-2 intrinsics to lift depth to point maps. This is stated in the text, but the qualitative figure caption should repeat it, since the comparison may otherwise appear unfair or unfair in the opposite direction.
  5. [Sec. 5, Limitations] The limitations paragraph focuses on regression-based boundary sharpness. Given the evaluation concerns above, consider also acknowledging that the local fine-detail metric uses per-segment alignment and synthetic data.

Circularity Check

0 steps flagged

No significant circularity: the method is a constructive architecture change validated against external benchmarks and controlled ablations.

full rationale

The derivation chain is not circular. MoGe-3 builds on the publicly defined MoGe-2 base model and adds an SSR module whose residual updates are learned under an affine-invariant loss (Eq. 1) inherited from earlier work; no equation in the method reduces to its own target. The sparse voxel shell is constructed from the current point map, but the U-Net output is a learned log-depth residual (Eqs. 3-5), and the ablations (2D Conv vs. SSR, D sweep, feature injection) provide independent evidence for the architectural claims. Quantitative claims are checked against external datasets (NYUv2, KITTI, ETH3D, etc.) and against prior models including the authors' own MoGe-2. The per-segment alignment in Appendix B.2 is an explicitly disclosed evaluation choice applied equally to all baselines, so it is a validity/generalizability concern rather than a circular reduction; it does not force the method's margin over other methods, since within-segment shape errors remain measurable under the same protocol. The Limitations section candidly notes regression-based boundary ambiguity, further supporting that the paper does not hide its weaknesses. Therefore, no load-bearing step reduces by construction to its input.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical or mathematical entities are introduced. The sparse voxel shell is a computational representation with exactly HW occupied voxels (one per pixel), not an independent entity with falsifiable content; no new forces, fields, or conserved quantities are postulated. The ledger's free parameters are concentrated in the voxelization resolution D, the iteration schedule, the evaluation-protocol thresholds, and training loss hyperparameters.

free parameters (4)
  • Voxel resolution D (log-depth quantization) = D = 200
    Depth-axis discretization of the voxel shell (Eq. 3). Swept 100–400 in Sec. 4.5 / Table 3; D=400 degrades both global and local metrics and D=200 is chosen as the trade-off. All SSR behavior depends on this hand-set granularity.
  • Refinement iterations K = K = 3 at training; applied up to 7 at inference
    Number of self-guided update cycles (Eq. 4). Fig. 4 shows metrics improve through K=5–7, so the choice is robust, but it is a hand-set schedule rather than an emergent property.
  • Local fine-detail metric thresholds (evaluation protocol) = 3.0·σ_MAD; element sizes {3,5,9,17}; ρ≥0.3; |Si|≥5; |Si|≤0.05HW; δ 0.01 / 1.01
    Appendix B parameters of the mask pipeline and per-segment shift alignment that define the headline local metric. Chosen by the authors; the central 'significantly outperforms' claim is measured on this instrument.
  • Loss weights and local partition scales = (λg,λl,λe,λm,λn,λs) = (1.0,1.0,1.0,0.1,0.1,0.1); α ∈ {4,16,64}
    Appendix A.2 hyperparameters, including the radial-partition granularity (Eq. 10). Standard training choices; they affect behavior but are not fitted to test data.
axioms (4)
  • domain assumption The architectural-mismatch hypothesis: 2D decoding of 3D geometry mixes features across depth discontinuities and causes the observed fine-detail distortions.
    Sec. 1 motivation; the voxelization design follows from this hypothesis. Partially supported by the 2D-conv ablation (Table 3) but not independently proven as the cause.
  • domain assumption Pixel-accurate synthetic ground truth transfers to real-world fine-detail geometry.
    Sec. 3.3: the refiner receives gradients only from synthetic samples, and local metrics are evaluated only on synthetic Spring/Synth4K. Real-world fine-detail benefit is assumed, not measured.
  • domain assumption Fixing image coordinates (u,v) and refining only log-depth is sufficient to correct the observed distortions.
    Eq. 2 and Sec. 3.2.3: 'the image-axis coordinates (u,v) remain fixed throughout the process'. Lateral (X,Y) errors are not corrected; if distortions like 'twisted fences' include lateral misplacement, a depth-only residual is incomplete.
  • domain assumption Per-segment translation alignment with a shared global scale is a fair measure of local geometric fidelity.
    App. B.2, Eqs. 17–18: each segment's prediction is shifted to best match GT before computing Rel/δ0.01. This lenient alignment can hide per-object depth or translation errors, and its use as a measure of '3D geometric quality' (Sec. 4.3) is an assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 22825 in / 23962 out tokens · 201922 ms · 2026-08-01T16:30:41.627141+00:00 · methodology

0 comments
read the original abstract

Monocular geometry estimation has recently achieved impressive performance across diverse scenes. However, state-of-the-art models still face notable distortion in local 3D structure, especially in fine details, like thin structures and small objects. We attribute this limitation to an architectural mismatch: most current models decode 3D geometry within a 2D parameterization, where feature interactions are governed by image-plane proximity rather than true 3D spatial relationships. This inadvertently mixes features from geometrically distant surfaces, resulting in over-smoothed geometry particularly around thin or elongated structure. In this paper, we propose MoGe-3, a fine-detail monocular geometry estimation model with Self-Guided Sparse 3D Refinement (SSR) that lifts monocular geometry modeling from 2D image space to 3D space for high-fidelity metric-scale point maps. MoGe-3 lifts the coarse point map from a foundation base model onto a sparse voxel shell and refines it via SSR. The SSR employs sparse convolutions that aggregate features based on 3D spatial locality, avoiding feature mixing across depth discontinuities. Extensive experiments on diverse datasets demonstrate that MoGe-3 significantly outperforms existing approaches in recovering fine detailed 3D geometry across both quantitative metrics and qualitative visualizations. Project page: https://qft-333.github.io/moge3page/

Figures

Figures reproduced from arXiv: 2607.17967 by Chengtang Yao, Jianfeng Xiang, Jiaolong Yang, Lingyu Kong, Ruicheng Li, Ruicheng Wang, Sicheng Xu.

Figure 1
Figure 1. Figure 1: Our method produces geometrically undistorted and detail-preserving metric scale point maps from a single image. Compared to prior approaches [45, 22, 58], it better preserves fine structures and recovers more accurate local geometry, leading to improved 3D fidelity in challenging cases. Abstract Monocular geometry estimation has recently achieved impressive performance across diverse scenes. However, stat… view at source ↗
Figure 2
Figure 2. Figure 2: Method overview. (a) The Base model provides an initial point map and 2D features. (b) The SSR module iteratively refines this geometry: current estimates are discretized into sparse voxel shells that guide 3D inference, which then updates the point map for the subsequent cycle. (c) The Sparse 3D U-Net employs multi-scale sparse 3D convolutions, injecting 2D features from base model, and predicts log-depth… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of point maps and disparity maps on unseen test images. Since [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of refinement iterations. Left: point map visualizations demonstrate progressively [PITH_FULL_IMAGE:figures/full_fig_p009_4.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

65 extracted references · 33 linked inside Pith

  1. [1]

    Apollo synthetic dataset, 2019

    Baidu Apollo. Apollo synthetic dataset, 2019. URL https://developer.apollo.auto/ synthetic.html. Accessed: 2025-03-06

    2019

  2. [2]

    ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data

    Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (...

    2021

  3. [3]

    Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. URL https://arxiv.org/ abs/2302.12288

    Pith/arXiv arXiv 2023

  4. [4]

    Richter, and Vladlen Koltun

    Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second,

  5. [5]

    D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor,European Conf. on Computer Vision (ECCV), Part IV , LNCS 7577, pages 611–625. Springer-Verlag, October 2012. 10

    2012

  6. [6]

    4d spatio-temporal convnets: Minkowski convolutional neural networks, 2019

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks, 2019. URL https://arxiv.org/abs/1904. 08755

    2019

  7. [7]

    Objaverse: A universe of annotated 3d objects, 2022

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects, 2022. URLhttps://arxiv.org/abs/2212.08051

    Pith/arXiv arXiv 2022

  8. [8]

    Mid-air: A multi-modal dataset for extremely low altitude drone flights

    Michael Fonder and Marc Van Droogenbroeck. Mid-air: A multi-modal dataset for extremely low altitude drone flights. InConference on Computer Vision and Pattern Recognition Workshop (CVPRW), June 2019

    2019

  9. [9]

    Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image, 2024

    Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image, 2024. URLhttps://arxiv.org/abs/2403.12013

    Pith/arXiv arXiv 2024

  10. [10]

    Jakob Geyer, Yohannes Kassahun, Mentar Mahmudi, Xavier Ricou, Rupesh Durgesh, Andrew S. Chung, Lorenz Hauswald, Viet Hoang Pham, Maximilian Mühlegg, Sebastian Dorn, Tiffany Fernandez, Martin Jänicke, Sudesh Mirashi, Chiragkumar Savani, Martin Sturm, Oleksandr V orobiov, Martin Oelker, Sebastian Garreis, and Peter Schuberth. A2D2: Audi Autonomous Driving D...

    2020

  11. [11]

    3d semantic segmentation with submanifold sparse convolutional networks, 2017

    Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks, 2017. URL https://arxiv.org/abs/ 1711.10275

    Pith/arXiv arXiv 2017

  12. [12]

    Gómez, Manuel Silva, Antonio Seoane, Agnés Borràs, Mario Noriega, German Ros, Jose A

    Jose L. Gómez, Manuel Silva, Antonio Seoane, Agnés Borràs, Mario Noriega, German Ros, Jose A. Iglesias-Guitian, and Antonio M. López. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes.Neurocomputing, 637:130038, 2025. ISSN 0925-2312. doi: 10.1016/j.neucom.2025.130038. URL http://dx.doi.org/10.1016/j. neucom.20...

    arXiv 2025

  13. [13]

    Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10579–10596, December 2024. ISSN 1939-3539. doi...

    arXiv 2024

  14. [14]

    Deep- mvs: Learning multi-view stereopsis

    Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deep- mvs: Learning multi-view stereopsis. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  15. [15]

    On the importance of accurate geometry data for dense 3d vision tasks

    HyunJun Jung, Patrick Ruhkamp, Guangyao Zhai, Nikolas Brasch, Yitong Li, Yannick Verdie, Jifei Song, Yiren Zhou, Anil Armagan, Slobodan Ilic, et al. On the importance of accurate geometry data for dense 3d vision tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 780–791, 2023

    2023

  16. [16]

    Repurposing diffusion-based image generators for monocular depth estimation,

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Kon- rad Schindler. Repurposing diffusion-based image generators for monocular depth estimation,

  17. [17]

    Comparison of monocular depth estimation methods using geometrically relevant metrics on the ibims-1 dataset.Computer Vision and Image Understanding (CVIU), 191:102877, 2020

    Tobias Koch, Lukas Liebel, Marco Körner, and Friedrich Fraundorfer. Comparison of monocular depth estimation methods using geometrically relevant metrics on the ibims-1 dataset.Computer Vision and Image Understanding (CVIU), 191:102877, 2020. doi: 10.1016/j.cviu.2019.102877

    arXiv 2020

  18. [19]

    EDEN: Multimodal Synthetic Dataset of Enclosed garDEN Scenes

    Hoang-An Le, Partha Das, Thomas Mensink, Sezer Karaoglu, and Theo Gevers. EDEN: Multimodal Synthetic Dataset of Enclosed garDEN Scenes. InProceedings of the IEEE/CVF Winter Conference of Applications on Computer Vision (WACV), 2021. 11

    2021

  19. [20]

    Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond, 2023

    Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond, 2023. URL https://arxiv.org/abs/2309.16553

    Pith/arXiv arXiv 2023

  20. [21]

    Megadepth: Learning single-view depth prediction from internet photos, 2018

    Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos, 2018. URLhttps://arxiv.org/abs/1804.00607

    Pith/arXiv arXiv 2018

  21. [22]

    Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Junhao Liew, Donny Y . Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views, 2025. URL https://arxiv.org/abs/2511.10647

    Pith/arXiv arXiv 2025

  22. [23]

    Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo

    Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andrés Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  23. [24]

    Indoor segmentation and support inference from rgbd images

    Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. InECCV, 2012

    2012

  24. [25]

    3d ken burns effect from a single image,

    Simon Niklaus, Long Mai, Jimei Yang, and Feng Liu. 3d ken burns effect from a single image,

  25. [26]

    Dinov2: Learning robust visual features without supervision, 2024

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick La...

    Pith/arXiv arXiv 2024

  26. [27]

    Unidepth: Universal monocular metric depth estimation, 2024

    Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation, 2024. URL https: //arxiv.org/abs/2403.18913

    Pith/arXiv arXiv 2024

  27. [28]

    Unik3d: Universal camera monocular 3d estimation, 2025

    Luigi Piccinelli, Christos Sakaridis, Mattia Segu, Yung-Hsu Yang, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unik3d: Universal camera monocular 3d estimation, 2025. URL https: //arxiv.org/abs/2503.16591

    Pith/arXiv arXiv 2025

  28. [29]

    URLhttps://arxiv.org/abs/1909.05483

    Pith/arXiv arXiv 1909

  29. [30]

    Vision transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. CoRR, abs/2103.13413, 2021. URLhttps://arxiv.org/abs/2103.13413

    Pith/arXiv arXiv 2021

  30. [31]

    Sam 2: Segment anything in images and videos, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos, 2024. URL https: //arxiv.org...

    Pith/arXiv arXiv 2024

  31. [32]

    3dvnet: Multi-view depth prediction and volumetric refinement

    Alexander Rich, Noah Stier, Pradeep Sen, and Tobias Höllerer. 3dvnet: Multi-view depth prediction and volumetric refinement. In2021 International Conference on 3D Vision (3DV), pages 700–709. IEEE, 2021

    2021

  32. [33]

    Unidepthv2: Universal monocular metric depth estimation made simpler

    Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Unidepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025

    Pith/arXiv arXiv 2025

  33. [34]

    German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M. Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    2016

  34. [35]

    BAD SLAM: Bundle adjusted direct RGB-D SLAM

    Thomas Schöps, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. InConference on Computer Vision and Pattern Recognition (CVPR), 2019. 12

    2019

  35. [36]

    Oriane Siméoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julie...

    Pith/arXiv arXiv 2025

  36. [37]

    Susskind

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding, 2021. URLhttps://arxiv.org/abs/2011.02523

    Pith/arXiv arXiv 2021

  37. [38]

    Searching efficient 3d architectures with sparse point-voxel convolution, 2020

    Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d architectures with sparse point-voxel convolution, 2020. URL https: //arxiv.org/abs/2007.16100

    Pith/arXiv arXiv 2020

  38. [39]

    Smd-nets: Stereo mixture density networks

    Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. InConference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  39. [40]

    Sparsity invariant cnns

    Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. InInternational Conference on 3D Vision (3DV), 2017

    2017

  40. [41]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception...

    2020

  41. [42]

    Flow-motion and depth network for monocular stereo and beyond.CoRR, abs/1909.05452, 2019

    Kaixuan Wang and Shaojie Shen. Flow-motion and depth network for monocular stereo and beyond.CoRR, abs/1909.05452, 2019. URLhttp://arxiv.org/abs/1909.05452

    Pith/arXiv arXiv 1909

  42. [43]

    Irs: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation, 2021

    Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. Irs: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation, 2021. URLhttps://arxiv.org/abs/1912.09678

    Pith/arXiv arXiv 2021

  43. [44]

    Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2025

    Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2025. URLhttps://arxiv.org/abs/2410.19115

    Pith/arXiv arXiv 2025

  44. [46]

    URLhttp://arxiv.org/abs/1908.00463

    Pith/arXiv arXiv 1908

  45. [47]

    Tartanair: A dataset to push the limits of visual slam,

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam,

  46. [48]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. InProceedings of the Neural Information Processing Systems Track on D...

    2021

  47. [49]

    Synscapes: A photorealistic synthetic dataset for street scene parsing.CoRR, abs/1810.08705, 2018

    Magnus Wrenninge and Jonas Unger. Synscapes: A photorealistic synthetic dataset for street scene parsing.CoRR, abs/1810.08705, 2018. URLhttp://arxiv.org/abs/1810.08705

    Pith/arXiv arXiv 2018

  48. [50]

    Moge-2: Accurate monocular geometry with metric scale and sharp details, 2025

    Ruicheng Wang, Sicheng Xu, Yue Dong, Yu Deng, Jianfeng Xiang, Zelong Lv, Guangzhong Sun, Xin Tong, and Jiaolong Yang. Moge-2: Accurate monocular geometry with metric scale and sharp details, 2025. URLhttps://arxiv.org/abs/2507.02546

    Pith/arXiv arXiv 2025

  49. [51]

    Sparse convolutional networks for surface reconstruction from noisy point clouds

    Tao Wang, Jing Wu, Ze Ji, and Yu-Kun Lai. Sparse convolutional networks for surface reconstruction from noisy point clouds. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3212–3221, 2024

    2024

  50. [52]

    Second: Sparsely embedded convolutional detection.Sensors, 18(10), 2018

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection.Sensors, 18(10), 2018. ISSN 1424-8220. doi: 10.3390/s18103337. URL https://www.mdpi.com/ 1424-8220/18/10/3337

    doi:10.3390/s18103337 2018

  51. [53]

    Depth anything: Unleashing the power of large-scale unlabeled data, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024. URL https: //arxiv.org/abs/2401.10891

    Pith/arXiv arXiv 2024

  52. [54]

    Depth anything v2, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. URLhttps://arxiv.org/abs/2406.09414

    Pith/arXiv arXiv 2024

  53. [55]

    Blendedmvs: A large-scale dataset for generalized multi-view stereo networks, 2020

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks, 2020. URLhttps://arxiv.org/abs/1911.10127

    Pith/arXiv arXiv 2020

  54. [56]

    Native and compact structured latents for 3d generation.arXiv preprint arXiv:2512.14692, 2025

    Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, et al. Native and compact structured latents for 3d generation.arXiv preprint arXiv:2512.14692, 2025. 13

    Pith/arXiv arXiv 2025

  55. [57]

    Pixel-perfect depth with semantics-prompted diffusion transformers, 2025

    Gangwei Xu, Haotong Lin, Hongcheng Luo, Xianqi Wang, Jingfeng Yao, Lianghui Zhu, Yuechuan Pu, Cheng Chi, Haiyang Sun, Bing Wang, Guang Chen, Hangjun Ye, Sida Peng, and Xin Yang. Pixel-perfect depth with semantics-prompted diffusion transformers, 2025. URL https://arxiv.org/abs/2510.07316

    arXiv 2025

  56. [58]

    Infinidepth: Arbitrary-resolution and fine-grained depth estimation with neural implicit fields, 2026

    Hao Yu, Haotong Lin, Jiawei Wang, Jiaxin Li, Yida Wang, Xueyang Zhang, Yue Wang, Xiaowei Zhou, Ruizhen Hu, and Sida Peng. Infinidepth: Arbitrary-resolution and fine-grained depth estimation with neural implicit fields, 2026. URLhttps://arxiv.org/abs/2601.03252

    arXiv 2026

  57. [59]

    Taskonomy: Disentangling task transfer learning, 2018

    Amir Zamir, Alexander Sax, William Shen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning, 2018. URL https://arxiv.org/ abs/1804.08328

    Pith/arXiv arXiv 2018

  58. [60]

    Structured3d: A large photo-realistic dataset for structured 3d modeling

    Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. InEuropean Conference on Computer Vision (ECCV), 2020

    2020

  59. [61]

    Omniworld: A multi-domain and multi-modal dataset for 4d world modeling, 2025

    Yang Zhou, Yifan Wang, Jianjun Zhou, Wenzheng Chang, Haoyu Guo, Zizun Li, Kaijing Ma, Xinyue Li, Yating Wang, Haoyi Zhu, Mingyu Liu, Dingning Liu, Jiange Yang, Zhoujie Fu, Junyi Chen, Chunhua Shen, Jiangmiao Pang, Kaipeng Zhang, and Tong He. Omniworld: A multi-domain and multi-modal dataset for 4d world modeling, 2025. URL https://arxiv. org/abs/2509.1220...

    arXiv 2025

  60. [62]

    Scannet++: A high-fidelity dataset of 3d indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InProceedings of the International Conference on Computer Vision (ICCV), 2023

    2023

  61. [63]

    Metric3d: Towards zero-shot metric 3d prediction from a single image, 2023

    Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image, 2023. URLhttps://arxiv.org/abs/2307.10984

    Pith/arXiv arXiv 2023

  62. [2019]

    URLhttp://arxiv.org/abs/1907.01341

    Pith/arXiv arXiv 1907

  63. [2020]

    URLhttps://arxiv.org/abs/2003.14338

    Pith/arXiv arXiv 2003

  64. [2024]

    URLhttps://arxiv.org/abs/2312.02145

    Pith/arXiv arXiv

  65. [2025]

    URLhttps://arxiv.org/abs/2410.02073

    Pith/arXiv arXiv