Dens3R: A Foundation Model for 3D Geometry Prediction

Chengfei Lyu; Jiangjing Lyu; Jingnan Gao; Qiaomu Ren; Xianze Fang; Xiaokang Yang; Xingyu Ren; Yichao Yan; Zhe Wang; Zhonglei Yang

REVIEW 5 major objections 7 minor 7 cited by

Dens3R: A Foundation Model for 3D Geometry Prediction

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Dens3R claims that building normals into the pointmap makes pointmaps, depth, normals, and matches jointly accurate and consistent from unposed images.

desk verdict Solid extension of the DUSt3R line with a real two-stage idea, but the load-bearing claim that normal supervision improves the pointmap itself is asserted rather than measured. read the letter →

arxiv 2507.16290 v1 pith:NO3SX6V3 submitted 2025-07-22 cs.CV

Xianze Fang , Jingnan Gao , Zhe Wang , Zhuo Chen , Xingyu Ren , Jiangjing Lyu , Qiaomu Ren , Zhonglei Yang

show 3 more authors

Xiaokang Yang Yichao Yan Chengfei Lyu

This is my paper · ORCID

classification cs.CV

keywords intrinsic-invariantpointmapvisualfoundationmodel3Dgeometrypredictionsurfacenormalestimationdepthimagematchinghigh-resolutioninferencetwo-stagetraining

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes Dens3R, a feed-forward dense 3D foundation model that takes unposed images (one or many, at various resolutions) and outputs, in a single pass, 3D pointmaps, depth maps, surface normals, camera parameters, and dense pixel matches. Its central claim is that these geometric quantities should not be learned separately: if the pointmap representation is made intrinsically invariant, so that the same surface carries consistent normal information across viewpoints, then normals, depth, pointmaps, and matching all become more accurate and mutually consistent. Dens3R trains in two stages, first a scale-invariant pointmap from cross-view matching losses, then an intrinsic-invariant pointmap with normal supervision, and it adds position-interpolated rotary positional encoding to keep high-resolution inputs from degrading. If the claim holds, one general-purpose backbone can replace separate depth, normal, matching, and reconstruction models, and can be fine-tuned for broader tasks such as segmentation and surface reconstruction.

What carries the argument

The central object is the intrinsic-invariant pointmap, a pointmap whose per-pixel representation is concatenated with the view-space surface normal so the same 3D structure carries consistent normal information across viewpoints. The two-stage training curriculum is the mechanism that makes it work: Stage 1 learns a scale-invariant pointmap under local and global 3D regression losses plus the MASt3R pixel matching (infoNCE) loss; Stage 2 fine-tunes that foundation with a normal prediction head and a normal loss $L_n$ under one-to-one supervision. A second supporting mechanism is position-interpolated rotary positional encoding, which re-encodes position $m$ at the original frequency for a longer sequence by $R'(x,m)=R(x, mL/L')$, so high-resolution inputs are interpolated rather than extrapolated. Together these pieces are what the paper credits for simultaneous accuracy on pointmaps, depth, normals, and matches.

What would settle it

A concrete check: train Dens3R with the intrinsic-invariant normal stage and with that stage ablated, and compare pointmap and depth errors, not only normal angle, on NYUv2 and ScanNet; if depth and pointmap errors stay flat while only normal metrics improve, the paper's unified-representation claim would be contradicted.

Watch

Extended reading notes

Core claim

The paper's discovery is that surface normals are the missing coupling term in pointmap-based 3D regression. Earlier DUSt3R-style models predict pointmaps with cross-view losses but ignore normals, and the paper finds that normals derived from such scale-invariant pointmaps are inaccurate. Dens3R's second stage concatenates the normal into the pointmap, defining an intrinsic-invariant pointmap $P_i^n = P_i \oplus n$, and switches from many-to-one to one-to-one supervision so each viewpoint's normal is optimized independently. With this representation and a shared encoder-decoder transformer, Dens3R reports lower angular errors than regression- and diffusion-based normal estimators, higher matching AUC on ZEB, ScanNet-1500, and MegaDepth-1500 than MASt3R and VGGT, competitive or better depth errors than MoGe and VGGT, and lower reprojection error on the Map-free pose benchmark.

Load-bearing premise

The argument depends on the premise that feeding surface normals into the pointmap makes pointmaps and depth more accurate, not just the normal head; if normal supervision only sharpens normals, the unified-representation thesis reduces to a multi-head model.

Editorial extensions

If this is right

A single unposed image pair yields internally consistent depth, normals, pointmaps, and matches, so downstream systems no longer need to align outputs from separately trained geometric models.
Monocular normal estimation becomes less ambiguous: the paper reports the lowest mean angular error among compared normal estimators on NYUv2, ScanNet, IBims-1, Sintel, and DIODE-outdoor.
The same pointmap backbone supports camera pose estimation from one pair of unposed images, with lower reprojection error and higher pose AUC than DUSt3R, MASt3R, and VGGT on the Map-free benchmark.
High-resolution and multi-resolution inputs remain usable because position-interpolated RoPE avoids the extrapolation failure that makes prior pointmap models produce overlapping or inconsistent geometry at 2K.
Because the backbone is frozen and only new heads are trained, the representation transfers to segmentation and to surface reconstruction via normal supervision, pointing to a common geometry backbone for many tasks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The intrinsic-invariant idea need not stop at normals: if the pattern holds, other per-view intrinsic maps with weak cross-view ambiguity, such as albedo or material properties, could be concatenated into the same pointmap using the same two-stage curriculum.
A decisive test the paper leaves open is whether normal supervision improves pointmap and depth error, not just normal angle; the ablations in Table 4 report only normal metrics, so the cross-task benefit that motivates the unified representation is asserted rather than measured.
The two-stage schedule suggests a general principle for multi-output 3D regression: learn cross-view geometry first and add per-view intrinsic constraints later, because coupling them from the start is unstable; this ordering could guide other joint-prediction architectures.
The position-interpolated RoPE is a resolution-extension technique that should transfer to any dense prediction transformer whose training resolution bounds its operating range, such as optical flow or stereo networks.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Fang et al. have a genuinely new combination here: two-stage training that starts with a scale-invariant MASt3R-style pointmap and then fine-tunes with normal supervision toward an 'intrinsic-invariant' representation, plus RoPE position interpolation for high-res inputs and a shared encoder-decoder. The normal results are strong—they beat DSINE, GeoWizard, StableNormal on NYUv2, ScanNet, IBims, Sintel, DIODE-outdoor—and the matching results on ZEB/ScanNet-1500/MegaDepth-1500 are competitive or better than MASt3R. For a foundation-model paper in this line, that is real empirical content.

The soft spot is exactly what the stress-test note flags. The paper's central claim is that coupling normal information into the pointmap improves the pointmap itself (Sec. 1 and Eq. 9/11). But the ablation for intrinsic-invariant training (Tab. 4/7) reports only normal metrics. No pointmap accuracy or depth accuracy for w/o IIT vs. Ours. The only depth comparison (Tab. 3) doesn't give a consistent win over MoGe/VGGT—on NYUv2, Dens3R's REL is 0.042 vs. 0.035 MoGe and 0.038 VGGT. So the evidence is fully compatible with the weaker story: Stage 2 improves the normal head while the pointmap stays basically the Stage-1 model. That doesn't kill the paper—the strong normal and matching results stand on their own—but it means the 'structural coupling' contribution is currently asserted, not demonstrated.

Other issues are secondary. No code or weights, so nothing can be checked externally. The multi-view post-processing (Sec. 3.3) is described at the level of 'constructing and optimizing a dense correspondence network,' which is too vague to evaluate. The training data table is refreshingly detailed, though the loss table suggests real-world data (type C) omits normal supervision, which is fine but worth stating more explicitly. The limitation on thin structures (Fig. 14) is honest.

My bottom line: this deserves a serious referee. It's a plausible, well-motivated extension of a hot line of work, with strong normal and matching results and an architecture that others will want to build on. But the key mechanistic claim—normals improve pointmaps—needs a direct test. My recommendation: send to peer review with a request for code/weights, and require the IIT ablation to also report pointmap or depth error, not just normals. If the authors can show that, this becomes a meaningful contribution; if not, it's still a useful multi-head model but not the unified geometry paper it claims to be.

Referee Report

5 major / 7 minor

Summary. The paper proposes Dens3R, a feed-forward dense 3D foundation model that takes unposed single images or image pairs as input and jointly predicts pointmaps, depth maps, surface normals, camera parameters, and dense matching features. The architecture is a shared encoder-decoder dense visual transformer with position-interpolated rotary positional encoding to support high-resolution inputs. Training proceeds in two stages: Stage 1 learns a scale-invariant pointmap with MASt3R-style losses (local/global 3D regression, pointmap normal loss, pixel matching loss), and Stage 2 adds explicit normal-head supervision to form an 'intrinsic-invariant pointmap,' followed by a coarse-to-fine schedule from 512 to 1024 pixels. Task-specific heads are then fine-tuned with the backbone frozen, and a verbally described post-processing pipeline is used for multi-view inference. Experiments report strong results in normal estimation (NYUv2, ScanNet, IBims, Sintel, DIODE), dense matching (ZEB, ScanNet-1500, MegaDepth-1500), and camera pose estimation (Map-free), plus competitive but not uniformly best monocular depth results (e.g., NYUv2 REL 0.042 versus 0.035 for MoGe).

Significance. The strengths of the paper are real: the headline results are held-out benchmark numbers rather than quantities derived from the model's own equations, so the main empirical claims are not circular; the normal-estimation gains (NYUv2 mean 16.1 versus 17.5 for Lotus-G; IBims 16.0 versus 17.2 for StableNormal) and dense matching gains (ZEB mean 64.5 versus 59.9 for MASt3R) are consistent and sizeable; the dataset curation table (Tab. 5) is unusually detailed; and the demonstration of segmentation and surface-reconstruction uses of the frozen backbone strengthens the foundation-model framing. If the two-stage design genuinely makes normal supervision improve the pointmap and downstream geometric outputs, this would be a useful advance over single-task regression systems. However, the central cross-task coupling claim is currently supported only by normal-metric ablations (Sec. 4.3, Tab. 4 and Tab. 7) and qualitative figures, not by quantitative pointmap or depth ablations; the multi-view pipeline is likewise not quantitatively evaluated; and no code or weights are available. The paper is substantial but needs targeted evidence before the headline claim can be taken as established.

major comments (5)

[Sec. 4.3, Tab. 4 and Tab. 7] The central claim, stated in Sec. 1 ('introducing normal information during geometric prediction can significantly improve the accuracy of point maps') and embodied in the Stage-2 design (Sec. 3.2, Eq. 11), is that normal-based intrinsic-invariant training improves the pointmap representation itself. However, the only quantitative ablation of the intrinsic-invariant training reports normal metrics (mean, median, delta-11.25 degree) exclusively; no pointmap accuracy or depth accuracy is reported for the w/o IIT variant versus Ours. The depth comparison in Tab. 3 is also mixed: Dens3R's NYUv2 REL is 0.042 versus 0.038 for VGGT and 0.035 for MoGe. The current evidence is therefore consistent with the weaker interpretation that Stage 2 improves only the normal head, leaving the pointmap essentially at the Stage-1 level. Please add quantitative pointmap and depth metrics, such as depth REL/RMSE and a pointmap chamfer or F-score measure on NYUv2 and DIODE, for the w/o IIT, w/o C2F, and Ours variants.
[Sec. 4.2 and contribution list] The first listed contribution is pointmap reconstruction, and the title emphasizes 3D geometry prediction, yet pointmap quality is only ever shown qualitatively (Figs. 5-7 and 23); no quantitative pointmap metric is reported against DUSt3R, MASt3R, MoGe, or VGGT. Please add a quantitative pointmap evaluation (e.g., chamfer distance, F-score, or accuracy at thresholds) on a held-out set or an existing benchmark, paired with the same baselines used in Tab. 3.
[Sec. 3.3 multi-view inputs] The multi-view input pipeline is described only verbally: 'establish geometric mappings between different viewpoints by constructing and optimizing a dense correspondence network across views.' No equations, pseudocode, optimization objective, or hyperparameters are given, and the only evidence is qualitative (Fig. 23). Since multi-view geometric consistency is claimed as an outcome and the post-processing is a stated contribution, please specify the algorithm and include a quantitative evaluation showing that multi-view inputs improve consistency and accuracy relative to single-pair inference.
[Sec. 4.3 and Fig. 8a] The position-interpolated rotary positional encoding is a stated contribution (Sec. 3.1, Eq. 2), but Tab. 4 contains no row for w/o position interpolation, and Fig. 8a and Fig. 22 are qualitative. Please report a quantitative comparison with and without the position-interpolated RoPE at resolutions 512 and 1024 (for example, normal or matching metrics), and, if the claim that interpolation is more stable than direct extrapolation is retained, support it with a comparison against explicit extrapolation.
[Sec. 3.2, Eq. 9] The core 'intrinsic-invariant pointmap' is never formally defined. Section 3.2 describes a change from a 'one-to-many' mapping to a 'one-to-one' mapping and defines P^n_i = P_i + n (Eq. 9), but it is not specified what n is, what the concatenation produces, or what invariance the pointmap satisfies after Stage 2. In addition, Stage 1 already includes the Pointmap Normal Loss L_pts_n (Eq. 6 and Eq. 8), which is in tension with the narrative that normals enter only in Stage 2, and the assertion in Sec. 4.3 that jointly training pointmap and normal from the start is unstable is made without any supporting experiment. Please formalize the representation, disambiguate the roles of L_pts_n and L_n, and provide evidence (training curves or an ablation with the normal loss included from Stage 1) for the instability claim.

minor comments (7)

[Tab. 1 and Tab. 6] The DSINE row reports identical values for NYUv2 and ScanNet (Mean 18.6, Med 9.9, delta-11.25 degree 56.1) in both Tab. 1 and Tab. 6; please verify these entries.
[Tab. 2] Several cells in Tab. 2 are run together (e.g., the SIFT row contains '14.733.4 7.6'); please reformat the table for readability.
[Eq. 3 and Eq. 4] The normalization factors z_v and bar z_v are defined through garbled superscript notation in the text; the intended definition (average distance of valid points to the origin) should be written explicitly.
[Eq. 11] The pixel matching loss L_match present in Eq. 8 is absent from Eq. 11 without comment; please state whether the matching head is frozen in Stage 2 and why L_match is dropped.
[Tab. 3 footnote] The footnote reads 'we utilize Lotu-G disparity model'; this should be 'Lotus-G', and the asterisk footnotes in Tab. 1 and Tab. 3 should be defined consistently.
[References] MegaDepth is cited twice, as [Li and Snavely 2018a] and [Li and Snavely 2018b], with identical titles; please merge into one entry.
[Availability] The manuscript is silent on code and model weight availability; for a foundation-model paper, please add a statement, as this materially affects reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Dens3R's headline results are external benchmark evaluations, and no predicted quantity reduces to a fitted parameter or self-citation by construction.

full rationale

The paper's central claims are empirical: Dens3R is trained with the losses in Eqs. 4-11 and evaluated on external benchmarks (NYUv2, ScanNet, ZEB, MegaDepth-1500, Map-free, etc.). None of the reported metrics is defined as a fitted constant or as the model's own training target in disguised form. The two-stage intrinsic-invariant training is a design choice; the ablation in Sec. 4.3/Tab. 4/Tab. 7 reports only normal metrics, so the claim that normal supervision improves pointmap accuracy is under-supported, but that is a missing-evidence/correctness issue, not a circular derivation: Eq. 11 does not force the pointmap improvement to hold by construction, and the claim remains externally testable. The position-interpolated RoPE (Eq. 2) is adapted from external LLM context-extensions work (Chen et al. 2023), not from the authors' own prior results, and its benefit is shown through comparison with DUSt3R/VGGT at high resolution. The paper cites prior DUSt3R-family methods as baselines and building blocks, but those citations are external works and are not used to forbid alternatives or to import a uniqueness theorem. The Appendix I limitation about thin structures is an honest limitation statement and does not indicate circularity. Overall, the derivation chain is self-contained in the sense that the reported predictions are benchmark-measured outputs of a learned model, not analytical consequences of its own definitions.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

All numbers and assumptions above are taken from the paper's equations and tables. The loss weights and dataset proportions are the main free parameters; they are tuned on the same benchmark family, which raises a mild circularity note. No new physical entities are introduced; the intrinsic-invariant pointmap is a learned representation whose independent evidence is limited to the paper's own ablations.

free parameters (6)

Stage-1 loss weights = Set to 1.0, 0.1, and 0.075 for eta1, eta2, and eta3 in the paper.
Eq. (8); they balance pointmap, matching, and normal losses in stage 1.
Stage-2 loss weights = Set to 1.0, 0.1, and 1.0 for lambda1, lambda2, and lambda3 in the paper.
Eq. (11); they balance pointmap and normal losses in stage 2.
Matching loss temperature = Not reported in the paper.
Eq. (7); tau scales descriptor similarities in the infoNCE matching loss.
Training resolutions = Set to 512 pixels for the coarse stage and 1024 pixels for the fine stage.
Coarse-to-fine schedule in Sec. 3.2 and Sec. A; high-resolution data restricted to type A plus a minor portion of type B.
Position interpolation factor = Not reported in the paper.
Eq. (2); the ratio L/L' controls RoPE interpolation at high resolutions.
Dataset mixing ratios = Listed in Table 5 with percentages such as 6.77 and 13.53.
Authors state the proportions were carefully allocated for optimal training performance.

assumptions (6)

standard math RoPE interpolation formula and its smoothness assumptions (Eq. 2).
Interpolation of rotary positional encoding is adapted from LLM context extension (Chen et al. 2023).
domain assumption Geometric quantities are inherently interconvertible and joint regression improves accuracy.
Sec. 1 and Sec. 3 state this as the motivation; it is the thesis the model is built on.
domain assumption Surface normals provide intrinsic invariance that simplifies training.
Sec. 1 and Sec. 3.2 argue normals reduce monocular ambiguity and aid convergence.
ad hoc to paper Jointly training pointmaps and normals from stage 1 causes instability.
Sec. 4.3 asserts this, but no ablation comparing joint single-stage training against the two-stage schedule is provided.
domain assumption Pairwise camera poses used to transform normals in Eq. (6) are accurate enough.
Lpts_n in stage 1 transforms normals between camera frames using predicted geometry; inaccurate poses would corrupt supervision.
domain assumption Confidence loss in prior works causes models to ignore complex scenes.
Sec. 3.2 claims this to justify removing confidence weighting; no citation or experiment supports it.

invented entities (1)

Intrinsic-invariant pointmap
purpose: A pointmap representation augmented with normal information to align geometric predictions across viewpoints.
Sec. 3.2, Eq. (9) defines P_i plus n; its benefit is only shown via internal normal ablations (Tab. 4), not by external replication or downstream pointmap metrics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dens3R: A Foundation Model for 3D Geometry Prediction." pith.science (2026). https://pith.science/paper/NO3SX6V3

@misc{pith2026250716290,
  author       = {Pith},
  title        = {Pith review of: Dens3R: A Foundation Model for 3D Geometry Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NO3SX6V3}},
  note         = {Machine review of arXiv:2507.16290}
}

read the original abstract

Recent advances in dense 3D reconstruction have led to significant progress, yet achieving accurate unified geometric prediction remains a major challenge. Most existing methods are limited to predicting a single geometry quantity from input images. However, geometric quantities such as depth, surface normals, and point maps are inherently correlated, and estimating them in isolation often fails to ensure consistency, thereby limiting both accuracy and practical applicability. This motivates us to explore a unified framework that explicitly models the structural coupling among different geometric properties to enable joint regression. In this paper, we present Dens3R, a 3D foundation model designed for joint geometric dense prediction and adaptable to a wide range of downstream tasks. Dens3R adopts a two-stage training framework to progressively build a pointmap representation that is both generalizable and intrinsically invariant. Specifically, we design a lightweight shared encoder-decoder backbone and introduce position-interpolated rotary positional encoding to maintain expressive power while enhancing robustness to high-resolution inputs. By integrating image-pair matching features with intrinsic invariance modeling, Dens3R accurately regresses multiple geometric quantities such as surface normals and depth, achieving consistent geometry perception from single-view to multi-view inputs. Additionally, we propose a post-processing pipeline that supports geometrically consistent multi-view inference. Extensive experiments demonstrate the superior performance of Dens3R across various dense 3D prediction tasks and highlight its potential for broader applications.

Figures

Figures reproduced from arXiv: 2507.16290 by the authors.

**Figure 1.** Dens3R is a feed-forward visual foundation model that takes unposed images as input and outputs high-quality 3D pointmap with unified geometric [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Overview of Dens3R. We propose Dens3R, a dense visual transformer backbone featuring a shared encoder-decoder architecture and multiple [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Normal comparison. We demonstrate that the normal derived directly [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (21 more)

**Figure 4.** Figure 4: Qualitative comparison of normal prediction. Dens3R generates more accurate and detailed normal maps than previous methods. Our method is capable [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Qualitative comparison of depth maps and pointmaps. We compare our method with previous DUSt3R-based methods and demonstrate high-quality [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: High-quality geometric predictions for high-resolution (2K) inputs. Please zoom in to better observe the fine-grained details. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: High-quality unified geometric predictions for various scenarios. We demonstrate accurate normal and depth predictions with high-quality 3D [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

**Figure 8.** Figure 8: Ablation and downstream applications [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 13.** Figure 13: It can be seen that Dens3R can handle normal predictions [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

**Figure 9.** Figure 9: Additional visual results for unified geometric predictions of monocular inputs by our methods [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: Additional visual results for unified geometric predictions of 2-view images inputs by our method. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

**Figure 11.** Figure 11: Normal prediction comparison of different training stages. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

**Figure 12.** Figure 12: Normal comparison of Kitti dataset. We present more normal com [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

**Figure 13.** Figure 13: Normal comparison with MoGe. We provide more normal compar [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]

**Figure 14.** Figure 14: Limitations. Despite that our method outperforms previous methods [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]

**Figure 15.** Figure 15: More qualitative comparison of normal map. We provide more normal comparison of both object-centric and human scenes. Dens3R is able to produce [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

**Figure 16.** Figure 16: More qualitative comparison of normal map. We provide more normal comparison of both indoor and outdoor scenes. Dens3R is able to produce [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]

**Figure 17.** Figure 17: Additional depth comparison. We provide more depth comparison with previous methods and our method can predict more accurate and detailed [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]

**Figure 18.** Figure 18: Additional depth comparison with VGGT. We compare our depth prediction results with VGGT and Dens3R demonstrates more robust and accurate [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]

**Figure 19.** Figure 19: Additional depth comparison of indoor scenes with VGGT. Dens3R demonstrates more accurate results for human depth estimation. [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]

**Figure 20.** Figure 20: Additional depth comparison of outdoor scenes with VGGT. We compare our depth prediction results of autonomous driving dataset. Our methods [PITH_FULL_IMAGE:figures/full_fig_p022_20.png]

**Figure 21.** Figure 21: Image-matching visualization. We provide the visualization of our dense and accurate image-matching results. The accuracy is also presented on the [PITH_FULL_IMAGE:figures/full_fig_p023_21.png]

**Figure 22.** Figure 22: Additional high-resolution inference comparison. We provide more high-resolution inference results to demonstrate the effectiveness of the proposed [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]

**Figure 23.** Figure 23: Multi-view reconstruction results. We demonstrate high-quality 3D reconstruction for various scenarios without known camera poses. The predicted [PITH_FULL_IMAGE:figures/full_fig_p025_23.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

FoundationGeo: Learning Spatial Pixel-Wise Fields for Monocular Metric Geometry
cs.CV 2026-07 conditional novelty 6.0 of 10

Pixel-wise scale and ray-direction fields plus focal-diverse synthetic data convert a strong relative geometry backbone into the best average zero-shot monocular metric depth across seven benchmarks.
Towards Consistent Video Geometry Estimation
cs.CV 2026-05 unverdicted novelty 6.0 of 10

ViGeo is a feed-forward transformer for video geometry that introduces dynamic chunking attention and a completion-based data refinement framework to achieve SOTA on depth, normals, and point map estimation.
4DVGGT-D: 4D Visual Geometry Transformer with Improved Dynamic Depth Estimation
cs.CV 2026-05 unverdicted novelty 6.0 of 10

A training-free two-pass adaptation of VGGT, with attention-based motion masking and inverse-variance depth fusion, improves dynamic-scene point-cloud reconstruction on DyCheck.
2K Retrofit: Entropy-Guided Efficient Sparse Refinement for High-Resolution 3D Geometry Prediction
cs.CV 2026-03 conditional novelty 5.5 of 10

Entropy-guided sparse refinement upgrades frozen low-resolution geometric foundation models to accurate 2K depth and pointmap outputs at a fraction of full-resolution cost.
GPOcc++: Unified Sparse Gaussian Occupancy Prediction with Visual Geometry Priors
cs.CV 2026-07 conditional novelty 5.0 of 10

A unified framework converts surface geometry priors into sparse Gaussian occupancy predictions and extends it to multi-view and temporal inputs.
MoWorld: A Flash World Model
cs.CV 2026-07 conditional novelty 5.0 of 10

MoWorld achieves up to 50 FPS real-time interactive world simulation on NPUs by combining a 3D-native data engine, curriculum cross-frame pretraining, autoregressive distillation, and mixed-precision parallel inference.
WorldMirror: Universal 3D World Reconstruction with Any-Prior Prompting
cs.CV 2025-10 conditional novelty 5.0 of 10

A unified feed-forward network that ingests optional geometric priors and jointly predicts point maps, depth, camera poses, normals, and 3D Gaussians, reporting state-of-the-art results on multiple 3D benchmarks.

Reference graph

Works this paper leans on

11 extracted references · 7 canonical work pages · cited by 7 Pith papers

[4]

Wonder3D: Single Image to 3D Using Cross-Domain Diffusion. InCVPR. 9970–9980. Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai

work page
[5]

Ricardo Martin-Brualla, Noha Radwan, Mehdi S

Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering.CVPR (2024). Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. 2021. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. InCVPR. 7210–7219. Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nal...

work page 2024
[6]

Spring: A High-Resolution High-Detail Dataset and Benchmark for Scene Flow, Optical Flow and Stereo. InCVPR. 4981–4991. Iaroslav Melekhov, Aleksei Tiulpin, Torsten Sattler, Marc Pollefeys, Esa Rahtu, and Juho Kannala. 2019. DGC-Net: Dense Geometric Correspondence Network. InW ACV. 1034–1042. Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T...

arXiv 2019
[7]

Pattern Anal

Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero- Shot Cross-Dataset Transfer.IEEE Trans. Pattern Anal. Mach. Intell.44, 3 (2022), 1623–1637. Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick La- batut, and David Novotny. 2021. Common Objects in 3D: Large-Scale Learning and Evaluation of Real-Life 3D Category...

work page 2022
[10]

DUSt3R: Geometric 3D Vision Made Easy. InCVPR. 20697–20709. Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. 2020b. TartanAir: A Dataset to Push the Limits of Visual SLAM. (2020), 4909–4916. Xiaolong Wang, David F. Fouhey, and Abhinav Gupta. 2015. Designing deep networks for surface ...

work page 2020
[11]

Input-level Inductive Biases for 3D Reconstruction. InCVPR. 6166–6176. Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. 2023. Metric3D: Towards Zero-shot Metric 3D Prediction from A Single Image. InICCV. 9009–9019. Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip- Splatting...

work page 2024
[2016]

The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes. InCVPR. 3234–3243. Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich

work page
[2020]

SuperGlue: Learning Feature Matching With Graph Neural Networks. In CVPR. 4937–4946. Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. 2019. Habitat: A Platform for Embodied AI Research. InICCV. 9338–9346. Philipp Schröppel, Jan ...

arXiv 2024

Show all 11 references

[2022]

In ECCV, Vol

Map-free Visual Relocalization: Metric Pose Relative to a Single Image. In ECCV, Vol. 13661. 690–708. Gwangbin Bae and Andrew J. Davison. 2024. Rethinking Inductive Biases for Surface Normal Estimation. InCVPR. 9535–9545. Aayush Bansal, Bryan C. Russell, and Abhinav Gupta. 201...

2024
[2023]

Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields. InICCV. 19697– 19705. Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. 2021. ARKitScenes - A Diverse Real-World ...

2023 arXiv
[2024]

RoMa: Robust Dense Feature Matching. InCVPR. 19790–19800. Ufuk Efe, Kutalmis Gokalp Ince, and A. Aydin Alatan. 2021. DFM: A Performance Baseline for Deep Feature Matching. InCVPRW. 4284–4293. Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. 2021. Omnidata: A Scal...

2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [4]

Wonder3D: Single Image to 3D Using Cross-Domain Diffusion. InCVPR. 9970–9980. Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai

work page

[2] [5]

Ricardo Martin-Brualla, Noha Radwan, Mehdi S

Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering.CVPR (2024). Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. 2021. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. InCVPR. 7210–7219. Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nal...

work page 2024

[3] [6]

Spring: A High-Resolution High-Detail Dataset and Benchmark for Scene Flow, Optical Flow and Stereo. InCVPR. 4981–4991. Iaroslav Melekhov, Aleksei Tiulpin, Torsten Sattler, Marc Pollefeys, Esa Rahtu, and Juho Kannala. 2019. DGC-Net: Dense Geometric Correspondence Network. InW ACV. 1034–1042. Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T...

arXiv 2019

[4] [7]

Pattern Anal

Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero- Shot Cross-Dataset Transfer.IEEE Trans. Pattern Anal. Mach. Intell.44, 3 (2022), 1623–1637. Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick La- batut, and David Novotny. 2021. Common Objects in 3D: Large-Scale Learning and Evaluation of Real-Life 3D Category...

work page 2022

[5] [10]

DUSt3R: Geometric 3D Vision Made Easy. InCVPR. 20697–20709. Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. 2020b. TartanAir: A Dataset to Push the Limits of Visual SLAM. (2020), 4909–4916. Xiaolong Wang, David F. Fouhey, and Abhinav Gupta. 2015. Designing deep networks for surface ...

work page 2020

[6] [11]

Input-level Inductive Biases for 3D Reconstruction. InCVPR. 6166–6176. Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. 2023. Metric3D: Towards Zero-shot Metric 3D Prediction from A Single Image. InICCV. 9009–9019. Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip- Splatting...

work page 2024

[7] [2016]

The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes. InCVPR. 3234–3243. Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich

work page

[8] [2020]

SuperGlue: Learning Feature Matching With Graph Neural Networks. In CVPR. 4937–4946. Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. 2019. Habitat: A Platform for Embodied AI Research. InICCV. 9338–9346. Philipp Schröppel, Jan ...

arXiv 2024