Pith. sign in

REVIEW 4 major objections 6 minor 22 cited by

RoMa v2 claims a new state-of-the-art for dense feature matching, beating prior matchers such as its predecessor RoMa on accuracy while running 1.7× faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 21:18 UTC pith:HRVHDOQJ

load-bearing objection RoMa v2 is a strong, honest dense-matching paper whose headline accuracy claim is partly undercut by training/eval overlap on AerialMegaDepth and MapFree. the 4 major comments →

arxiv 2511.15706 v3 pith:HRVHDOQJ submitted 2025-11-19 cs.CV

RoMa v2: Harder Better Faster Denser Feature Matching

Johan Edstedt , David Nordstr\"om , Yushan Zhang , Georg B\"okman , Jonathan Astermark , Viktor Larsson , Anders Heyden , Fredrik Kahl
show 2 more authors
M{\aa}rten Wadenb\"ack Michael Felsberg
This is my paper
classification cs.CV
keywords dense feature matchingmulti-view transformerDINOv3predictive covariancerelative pose estimationtraining mixtureCUDA kernelsubpixel accuracy
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 dense matching—estimating a warp and confidence for every pixel between two images—can be made both more accurate and faster than current best methods, not by one big idea but by a systematic set of improvements: a multi-view transformer matcher on top of frozen DINOv3 features, a new matching loss, a curated mix of wide- and small-baseline training datasets, and a cheaper refinement stage. The claimed result is a matcher that outperforms all prior dense matchers on six dense-matching benchmarks and on the MegaDepth-1500 and ScanNet-1500 pose benchmarks, while being 1.7× faster than its predecessor RoMa. It also predicts a per-pixel precision matrix, and the paper shows this covariance can significantly improve downstream geometry refinement. If true, this makes dense matching a more practical component for visual localization, 3D reconstruction, and other geometry tasks.

Core claim

RoMa v2 claims to be the current most accurate dense feature matcher, with reported AUC@5 of 62.8 on MegaDepth-1500 and 33.6 on ScanNet-1500, and lower endpoint errors than RoMa and UFM on all six dense-matching benchmarks (e.g., AerialMegaDepth EPE drops from 25.05 to 4.12). The gains come from a matcher that uses frozen DINOv3 features as the matching backbone, a multi-view transformer with an auxiliary negative-log-likelihood loss to learn multi-view context, and a training mixture that balances wide-baseline and small-baseline datasets. The refinement stage is made faster and more memory-efficient with a custom CUDA kernel, and a per-pixel precision matrix is added as an extra output, wh

What carries the argument

The central mechanism is the two-stage matching-then-refinement pipeline. The coarse matcher tokenizes both images with a frozen DINOv3 ViT-L encoder, then applies a multi-view transformer that alternates between global and frame-wise attention, producing a similarity matrix from which 'match embeddings' are computed and fed to a DPT head that predicts warp and confidence at stride 4. A new auxiliary loss—a dense directional negative log-likelihood over the similarity matrix—replaces the Gaussian-process regression used in RoMa. The refiners are lightweight VGG19-based CNNs at strides 4, 2, 1 that predict residuals, confidence deltas, and, new in this work, a 2×2 precision matrix per pixel v

Load-bearing premise

The load-bearing premise is that the hand-chosen mixture of training datasets (with some datasets weighted 0.01) and the MVS/flow-derived ground truth transfer to the full distribution of real-world matching tasks; neither the mixture weights nor the ground-truth cleanliness is ablated or validated.

What would settle it

One concrete test: hold out a dataset that mixes wide- and small-baseline pairs with a different distribution than the training mix (e.g., indoor handheld video with heavy motion blur and texture-poor walls), and compare RoMa v2 against RoMa and UFM. If the accuracy gap narrows or reverses, the curated training mixture rather than the architecture would be the main driver. A second test: recompute AerialMegaDepth after filtering sky pixels from the ground-truth depth, and see if the reported EPE drop from 25.05 to 4.12 shrinks.

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

If this is right

  • If the accuracy claims hold, RoMa v2 becomes the strongest dense matcher for subpixel correspondence, feeding more reliable matches into SfM, visual localization, and pose estimation.
  • The 1.7× speedup and reduced memory (custom CUDA kernel) make dense matching practical at higher resolutions and on resource-constrained hardware.
  • The per-pixel precision matrix output gives downstream geometry pipelines a principled way to weight residuals, improving RANSAC and refinement (AUC@1 rises by ~20 points on Hypersim).
  • Training on a diverse mixture (including small-baseline and aerial datasets) transfers to fine-grained details and texture-poor regions, as shown qualitatively on FlyingThings3D and NuScenes.
  • The matcher's robustness to resolution changes (normalized RoPE, low-frequency position embeddings) means it can be run at varying input sizes without the fixed-resolution restriction of UFM.

Where Pith is reading between the lines

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

  • The hand-tuned training mixture is never ablated, so the paper does not isolate how much of the gain comes from data versus architecture; a controlled study of the mixture weights would test whether the 0.01 weights on driving datasets are load-bearing for the claimed autonomous-driving robustness.
  • The ground truth for AerialMegaDepth leaks spurious sky depth (Appendix G), and the model learns to put confidence in sky pixels; evaluating on a corrected version of that benchmark could lower the reported AerialMegaDepth gains.
  • The precision-matrix output could be applied beyond matching, e.g., to optical flow or stereo, where per-pixel uncertainty is also needed; the paper only demonstrates it on one geometry-refinement setup.
  • On WxBS, RoMa v2 is slightly below RoMa on extreme modality changes (IR-to-RGB), so the 'robust to extreme appearance changes' claim is partial; a matcher that also masters that subset would be a natural next step.

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 / 6 minor

Summary. RoMa v2 is a dense feature matching architecture that extends RoMa with a frozen DINOv3 encoder, a multi-view transformer coarse matcher with an auxiliary NLL loss, three lightweight refiners with a custom CUDA kernel for local correlation, and a per-pixel predictive 2x2 precision matrix. Training is decoupled into matcher-then-refiner stages and uses a hand-curated mixture of seven wide-baseline and three small-baseline datasets. The paper reports state-of-the-art results on MegaDepth-1500, ScanNet-1500, six dense matching benchmarks, and a new SatAst benchmark, and shows a 1.7x speedup over RoMa with similar memory. The predictive covariance is demonstrated to improve covariance-weighted refinement on Hypersim.

Significance. The engineering contributions are substantial: the proposed architecture combines the robustness of frozen DINOv3 features with the efficiency of a decoupled two-stage pipeline, and the custom CUDA kernel addresses a real memory bottleneck. The covariance output is a useful novel capability, and the experiments cover a broad set of benchmarks with code release. However, the central claim of being 'significantly more accurate' is currently supported mainly by dense-matching numbers on benchmarks that appear in the training mixture, and the paper itself identifies a ground-truth artifact in AerialMegaDepth that could inflate the reported gains. Until the split overlap is resolved and the artifact is masked out, the magnitude of the claimed accuracy advantage remains uncertain, though the method is likely competitive.

major comments (4)
  1. [Sec. 3.4 Table 3 / Sec. 4.2 Table 7] Table 3 lists AerialMD and Map-Free in the training mixture with sampling weight 1.0, yet Table 7 reports dense matching EPE on AerialMegaDepth and MapFree. The paper does not state whether the evaluation scenes are excluded from training. If they overlap, the large EPE drops (25.05→4.12 and 8.55→2.03) may reflect in-distribution overfitting rather than generalization. Please report the exact scene splits used for evaluation and, if possible, evaluate on a held-out subset of these datasets. This is load-bearing for the 'significantly more accurate' claim.
  2. [Appendix G (Bias in AerialMegaDepth)] Appendix G documents spurious sky depth estimates in AerialMegaDepth (Fig. 14) and shows that RoMa v2 sometimes predicts confidence in sky regions (Fig. 13). The text states this 'possibly leaks into the warp estimate of RoMa v2.' This is a direct acknowledgment that the large EPE improvement on AerialMegaDepth (Table 7) may be inflated: RoMa v2's erroneous sky predictions can reduce EPE against a GT that includes sky depth, while RoMa receives large errors for not predicting sky. To support the claim, the authors should recompute dense matching metrics after masking invalid (sky/non-consistent) pixels and report the difference. Without this, the headline 84% relative EPE reduction is not reliable.
  3. [Sec. 3.4 Table 3 / Fig. 8] The training mixture is a stated contribution (#3), but no ablation isolates its effect. The weights (1.0 for seven datasets, 0.5 for FlyingThings3D, 0.01 for UnrealStereo4k and VKITTI2) are presented without sensitivity analysis, and the claim that VKITTI2 with weight 0.01 improves autonomous-driving robustness is supported only by a single qualitative comparison (Fig. 8). Please provide a systematic ablation, e.g., training with MegaDepth only (RoMa's setting), removing each dataset, or changing weights, to disentangle the contribution of the mixture from the architectural and loss changes. This is necessary to validate the 'curated diverse training distribution' claim and to understand what drives the dense-matching gains.
  4. [Sec. 4.1-4.2 / Tables 4,5,9] The abstract's 'significantly more accurate' is not consistently supported by the standard pose benchmarks: MegaDepth-1500 AUC@5 improves by only +0.2 (62.8 vs 62.6, Table 4), ScanNet-1500 by +1.8 (33.6 vs 31.8, Table 5), and WxBS mAA@10px drops by 5.4 points (60.8 vs 55.4, Table 9). The claimed significance rests mostly on the dense-matching benchmarks, which are the ones with potential training overlap. The authors should provide confidence intervals or a statistical comparison of the pose results, and discuss the WxBS degradation in more detail (the current explanation of the IR-to-RGB subset is not quantified).
minor comments (6)
  1. [Sec. 4.1] The statement 'RoMa v2 consistently outperforms all prior matchers on both benchmarks' is too strong on ScanNet, where RoMa v2 (33.6) is equal to MASt3R (33.6) and slightly below VGGT (33.9) in AUC@5; suggest 'matches or outperforms'.
  2. [Sec. 4.4 / Sec. 5] The WxBS mAA drop from 60.8 to 55.4 is described as 'slightly less robust' in the Limitations; this is a ~9% relative drop, comparable in magnitude to the gain on MegaDepth. Please quantify or revise the wording.
  3. [Sec. 4.6 / Table 10] The baseline 'w/o Σ^-1' is not fully described; state whether it uses the same refinement without covariance weighting, and specify the RANSAC parameters for reproducibility.
  4. [Sec. 3.3 / Eq. (6)] The hierarchical summation of precision matrices across strides is not intuitively justified; a brief explanation of why this parameterization is valid would help.
  5. [Sec. 4.5] The SatAst benchmark uses visually selected pairs and excludes pairs where homography annotation fails; this selection may bias the benchmark. Report the number of excluded pairs and provide the annotation protocol.
  6. [Throughout] Minor typos/clarity: 'demostrated' in Sec. 1; the sentence in Sec. 3.5 'We find that the approach used in RoMa ... generalizes the best' appears incomplete.

Circularity Check

1 steps flagged

Partial circularity: headline dense-matching gains on AerialMegaDepth/MapFree are reported on datasets that also appear in the training mixture (weight 1) with no documented held-out split; core ablations and other benchmarks remain independent.

specific steps
  1. fitted input called prediction [Sec. 3.4 Table 3 vs Sec. 4.2 Tables 6–7; App. B/G]
    "We train RoMa v2 on a mix of wide and small baseline two-view datasets, a summary of which is presented in Table 3. ... We evaluate dense matching performance in Table 6 and Table 7 on a wide array of datasets ... Table 3 lists AerialMD and Map-Free at weight 1; Table 7 evaluates AerialMegaDepth and MapFree."

    The training mixture in Table 3 includes AerialMD and Map-Free with weight 1, and §4.2 reports dense-matching EPE/PCK on AerialMegaDepth and MapFree without stating any held-out evaluation split. These metrics directly mirror the warp quantities minimized by the training losses, so unless the benchmark pairs are disjoint scenes/images—which the paper does not document—the headline gains (e.g., AerialMegaDepth EPE 25.05→4.12, MapFree 8.55→2.03) measure fit to the training label distribution rather than out-of-distribution generalization. App. G even shows the model learned AerialMegaDepth's spurious sky depth, consistent with training/eval sharing the same label noise. The paper does document hold-outs elsewhere (Hypersim, TartanAir), making this omission salient.

full rationale

The paper's derivation chain is mostly an empirical engineering story rather than a formal derivation. The architectural/loss contributions (Secs. 3.2–3.3) are validated by ablations on held-out Hypersim scenes and by comparisons on standard external benchmarks (MegaDepth-1500, ScanNet-1500, WxBS, TA-WB with held-out scenes), so those parts are not circular. The many self-citations to RoMa/DKM/UFM are background methodology, not load-bearing uniqueness claims. The one partial circularity is the dense-matching evaluation: AerialMegaDepth and MapFree appear in the training mixture (Table 3, weight 1) with no documented disjoint split in §4.2, and the EPE/PCK metrics directly mirror the training warp losses; App. G's sky-depth bias confirms the model absorbed the AerialMegaDepth label distribution. If the benchmark pairs are not disjoint, the headline 84% EPE reduction on AerialMegaDepth reduces to in-distribution fit. The SatAst benchmark is author-created and visually selected, which weakens its evidential weight but is not a definitional loop.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The paper contributes an empirical model; its inputs are standard vision modules (DINOv3, VGG19, alternating attention), and its claims rest on supervised training over proxy labels. The free parameters are dominated by hand-tuned mixture weights, loss weights, and thresholds — none ablated in a published table. The domain axioms are typical for the field but load-bearing: the GT warps come from depth/flow consistency that the paper itself shows to be imperfect (Appendix G), and all benchmark numbers are single-run. No genuinely new physical or computational entity is introduced: the covariance output is a trained prediction with in-distribution validation, not an ad hoc explanatory construct.

free parameters (6)
  • dataset mixture weights = MegaDepth 1, AerialMD 1, BlendedMVS 1, Hypersim 1, TartanAir v2 1, MapFree 1, ScanNet++ v2 1, UnrealStereo4k 0.01, VKITT
    Table 3; hand-chosen sampling probabilities balancing wide- and small-baseline data. The claimed robustness across benchmarks depends on these; no ablation isolates the mixture's contribution.
  • EMA decay alpha = 0.999
    Section 3.3; 'found a decay factor of α=0.999 to work well empirically' — chosen to remove sub-pixel bias; the reported accuracy depends on it.
  • precision-loss training threshold = ‖r‖ < 8 px; softplus floor 10⁻⁶
    Section 3.3; covariance is trained only on covisible residuals below 8 px — residuals beyond 8 px are discarded, which sharpens the NLL fit.
  • loss weights = 10⁻² overlap (matcher); 10⁻² overlap + 10⁻³ precision (refiners)
    Eqs. 2 and 8; chosen constants balancing losses; the precision loss weight is the smallest, which matters for the covariance claim.
  • overlap/GT thresholds = depth consistency τ=0.05 (Eq. 18); cycle consistency 5·10⁻³ ≈1.6px (Eq. 19); dataset pair overlap thresholds 0.01–0.2
    Appendix B and C; GT covisibility and training-pair selection are thresholded by hand-tuned values.
  • matching temperature and position-encoding scale = τ=1/10, ω=1 (fixed, non-learnable)
    Appendix A (Eq. 12) and Section 3.5; τ follows RoMa; ω=1 replaces RoMa's learnable ω=8 — a deliberate design choice affecting resolution robustness.
axioms (5)
  • domain assumption MVS/graphics-derived depth maps and flow fields yield correct ground-truth warps and covisibility labels
    Eqs. 18–19; the entire training signal is supervision from depth consistency or cycle consistency. Appendix G shows this fails for AerialMegaDepth sky pixels (spurious depth leaks into model predictions, Figs. 13–14).
  • domain assumption Frozen DINOv3 features are sufficiently robust that a linear-probe comparison (Table 1) predicts matching robustness
    Section 3.2, Table 1; the architecture's robustness rests on the foundation model being OOD-robust without fine-tuning; the evidence is a single linear-probe experiment on MegaDepth.
  • domain assumption Evaluating on fixed benchmark splits without repeated runs or error bars is representative of model quality
    Section 4; all tables report single passes. With +0.2 AUC differences on MegaDepth-1500 (Table 4) this assumption is load-bearing.
  • ad hoc to paper Pairwise Gaussian residuals with a per-pixel 2×2 precision matrix capture the relevant matching uncertainty
    Section 3.3, Eq. 5; the predictive-covariance claim assumes the NLL of a zero-mean 2D Gaussian is well-calibrated; the 8-pixel training cutoff and Cholesky parameterization are introduced for stability.
  • domain assumption The backward covariance can stand in for the forward covariance when computing weighted Sampson errors
    Appendix E, Eq. 20; the downstream refinement experiment approximates Σ⁻¹(A) by sampling Σ⁻¹(B→A) at the warped location — an approximation stated but not validated.
invented entities (1)
  • per-pixel predictive precision matrix Σ⁻¹ (covariance output) independent evidence
    purpose: enables uncertainty-weighted refinement and RANSAC scoring (Section 4.6, Table 10)
    Not a new physical entity but a new predicted quantity; it has a falsifiable handle — the Table 10 downstream refinement experiment and qualitative blur response (Fig. 9) — though shown only in-distribution on HyperSim.

pith-pipeline@v1.3.0-alltime-deepseek · 16962 in / 20809 out tokens · 184164 ms · 2026-08-03T21:18:36.709452+00:00 · methodology

0 comments
read the original abstract

Dense feature matching aims to estimate all correspondences between two images of a 3D scene and has recently been established as the gold standard due to its high accuracy and robustness. However, existing dense matchers still fail or perform poorly for many hard real-world scenarios, and high-precision models are often slow, limiting their applicability. In this paper, we attack these weaknesses on a wide front through a series of systematic improvements that together yield a significantly better model. In particular, we construct a novel matching architecture and loss, which, combined with a curated diverse training distribution, enables our model to solve many complex matching tasks. We further make training faster through a decoupled two-stage matching-then-refinement pipeline, and at the same time, significantly reduce refinement memory usage through a custom CUDA kernel. Finally, we leverage the recent DINOv3 foundation model along with multiple other insights to make the model more robust and unbiased. In our extensive set of experiments, we show that the resulting novel matcher sets a new state-of-the-art, being significantly more accurate than its predecessors. Code is available at https://github.com/Parskatt/romav2

Figures

Figures reproduced from arXiv: 2511.15706 by Anders Heyden, David Nordstr\"om, Fredrik Kahl, Georg B\"okman, Johan Edstedt, Jonathan Astermark, M{\aa}rten Wadenb\"ack, Michael Felsberg, Viktor Larsson, Yushan Zhang.

Figure 1
Figure 1. Figure 1: Radar chart of performance on benchmarks. RoMa v2 outperforms previous dense matchers on a wide range of pose estimation and dense matching tasks. Further details on these experiments can be found in Section 4. rely on precise and trustworthy correspondences in order to function robustly. Traditionally, these methods relied on sparse matches established purely through descriptor simi￾larity. In the last co… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative results. RoMa v2 excels at matching in diverse scenarios. We show a snapshot of results from different benchmarks. Below each image pair we visualize the dense warp by coloring each pixel by the RGB value from its estimated corresponding location in the opposite image. Brighter values mean lower warp confidence as output by the model. many challenging scenarios. For example, the recent RU￾BIK b… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of RoMa v2. We estimate bidirectional dense image warps W = {WA7→B ∈ R H×W×2 ,WB7→A ∈ R H×W×2 } and warp confidences p = {p A7→B ∈ R H×W×1 , p B7→A ∈ R H×W×1 } between two input images using a two-stage pipeline consisting of a matching and refinement stage. Different from recent SotA dense matchers, we additionally predict a precision matrix Σ−1 = {(Σ−1 ) A7→B ∈ R H×W×2×2 ,(Σ−1 ) B7→A ∈ R H×W×2×2… view at source ↗
Figure 4
Figure 4. Figure 4: Coarse matcher. We use a frozen DINOv3 feature ex￾tractor in the coarse matching stage. DINOv3 features from both input images are input to a Multi-view Transformer utilizing alter￾nating Attention. Dense Prediction Transformer (DPT) [30] heads output coarse warps W between the images and confidences p for 4x downsampled resolution [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Refiner internals. The coarse matcher predicts at a resolution 4x smaller than the original image size. The refiners output at the original resolution [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Subpixel bias of refinement. We observe that models exhibit subpixel fluctuations in their predictions throughout train￾ing, leading to bias. We propose a simple remedy through storing an exponential moving average (EMA). is additive in the precision parameterization to predict our final precision matrix as Σ −1 θi = X j≥i ∆Σ −1 θj . (6) We find empirically that our covariance improves perfor￾mance in down… view at source ↗
Figure 7
Figure 7. Figure 7: Finegrained objects in warp. Left: RoMa warp for small baseline pair. Note the missing warp for the guitar in the bottom right. Right: RoMa v2 warp. RoMa v2 is significantly better at capturing small objects with dynamic motion. The total refinement loss is Lrefiners = X i∈{1,2,4} Lwarp(rθi , pGT) + 10−2Loverlap(pθi , pGT) (8) + 10−3Lprecision(Σ −1 θi , detach(rθi )). 3.4. Data We train RoMa v2 on a mix of… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative example of covariances. We plot the pre￾dicted covariance for randomly sampled keypoints. In the right image, we have applied a linear kernel to simulate motion blur, yielding larger covariances, especially in the blur direction. residuals. First, only as post-processing, refining the out￾put of a classic point-based RANSAC. Second, we compare with using it to reweight the residuals used for sc… view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of training batch. Our data mixures is diverse and challenging, covering many types of scenes. Since RoMa v2 predicts only the forward covariance (and the residuals are two-sided), we approximate the full 4 × 4 covariance matrix of the matches by a block diagonal matrix where for each drawn correspondence pair x A, xB the covariance of points in I A are approximated by sam￾pling the backward… view at source ↗
Figure 11
Figure 11. Figure 11: Annotation of correspondences for SatAst: 1) We annotate four initial approximate correspondences. 2) We warp the satellite image using the homography obtained from the previous step and annotate ten accurate correspondences. 3) Visualization of the warp obtained by estimating a homography from the ten accurate annotations. 4) The ten accurate correspondences visualized in the original images, where we sc… view at source ↗
Figure 12
Figure 12. Figure 12: Accuracy of annotations on SatAst: A histogram over the reprojection errors of the 390 annotated correspondences in SatAst according to homographies estimated from the ten annno￾tations in each image. The image resolution of the satellite images is 3072 × 3072, so an error of 10 pixels is around 0.3% of the image width. F. Relative Pose Estimation VGGT: For evaluating VGGT on MegaDepth-1500, we follow the… view at source ↗
Figure 14
Figure 14. Figure 14: Spurious depth estimates in AerialMegaDepth. Depth from the scene leaks into the sky, causing some skypix￾els to be multi-view consistent. This possibly leaks into the warp estimate of RoMa v2. 5 [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 22 Pith papers

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

  1. One Video, One World: Turning Monocular Video into Physical 4D Scenes

    cs.CV 2026-06 unverdicted novelty 8.0

    OVOW reconstructs instance-level, simulation-ready 4D mesh scenes from monocular video via a four-stage training-free pipeline and introduces a new benchmark for structured Video-to-4D evaluation.

  2. REDI-Match: Rotation-Equivariant Distillation for Efficient and Robust Dense Matching

    cs.CV 2026-06 unverdicted novelty 7.0

    REDI-Match proposes rotation-equivariant distillation to transfer VFM semantics into a strictly equivariant encoder plus an entropy-driven alignment decoder, reporting SOTA accuracy and 1.9x speed on dense matching be...

  3. REDI-Match: Rotation-Equivariant Distillation for Efficient and Robust Dense Matching

    cs.CV 2026-06 unverdicted novelty 7.0

    REDI-Match uses rotation-equivariant distillation to transfer VFM semantics into a strictly equivariant encoder plus an entropy-driven alignment module, claiming SOTA accuracy and 1.9x speed on rotation-heavy benchmarks.

  4. Dynamic Inverse Rendering for Enhanced Material-Lighting Decomposition

    cs.CV 2026-07 conditional novelty 6.5

    Hand-held rigid motion yields significantly more accurate albedo and relighting than static multiview capture under the same far-field illumination, both synthetically and on real RGB video.

  5. FillGS: Filling Observation Gaps in 4D Gaussian Splatting via Viewpoint-Time Selection and Generative Refinement

    cs.CV 2026-07 conditional novelty 6.0

    FillGS actively selects spatiotemporal virtual viewpoints using rendering sensitivity and motion-aware observation density, then fine-tunes 4D Gaussian Splatting with reliability-masked generated images, improving spa...

  6. VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion

    cs.CV 2026-07 accept novelty 6.0

    VidMap recovers more accurate metric poses and calibration from long unconstrained videos than SOTA SLAM and SfM by treating temporal order as first-class and injecting metric depth into global optimization.

  7. DGSfM: Depth-Guided Scale-Aware Global Structure-from-Motion

    cs.CV 2026-07 accept novelty 6.0

    Monocular depth priors turn scale-ambiguous global SfM into a scale-aware pipeline that measurably improves camera pose accuracy on ETH3D and IMC2021.

  8. Glob3R: Global Structure-from-Motion with 3D Foundation Models

    cs.CV 2026-07 conditional novelty 6.0

    A frozen Pi3X backbone plus dense warping tracks and keyframe sliding-window global optimization yields more accurate, scalable SfM than feed-forward or classical baselines alone.

  9. OrthoTrack: Continuous 6-DoF UAV Trajectory Estimation Anchored in Public Orthophotos

    cs.CV 2026-06 unverdicted novelty 6.0

    OrthoTrack is a training-free system for continuous metric 6-DoF UAV pose estimation anchored in public orthophotos and surface models, with a new MovingDrone benchmark dataset.

  10. OrthoTrack: Continuous 6-DoF UAV Trajectory Estimation Anchored in Public Orthophotos

    cs.CV 2026-06 unverdicted novelty 6.0

    OrthoTrack is a training-free system for continuous 6-DoF UAV pose estimation anchored in public orthophotos and surface models, with a new MovingDrone benchmark dataset.

  11. Unified Motion-Action Modeling for Heterogeneous Robot Learning

    cs.RO 2026-06 unverdicted novelty 6.0

    UMA treats object motion and robot actions as co-evolving variables under a masked generative objective with hindsight relabeling and contrastive disentanglement to support multi-task pretraining and deployment across...

  12. SG2Loc: Sequential Visual Localization on 3D Scene Graphs

    cs.CV 2026-06 unverdicted novelty 6.0

    A particle-filter sequential localization method that matches per-patch semantic features from images to objects in a compact 3D scene graph via mesh projection and visibility.

  13. G2G: Exploiting Intra-Group Geometry for Inter-Group Pose Estimation

    cs.CV 2026-06 unverdicted novelty 6.0

    G2G attaches three small trainable modules to frozen backbones and reports state-of-the-art inter-group pose accuracy on four datasets spanning simulation, real cross-season, and sim-to-real transfer using only relati...

  14. Privacy-Preserving Structureless Visual Localization via Image Obfuscation

    cs.CV 2026-04 unverdicted novelty 6.0

    Simple image obfuscation enables privacy-preserving structureless visual localization with standard feature matchers and no pipeline changes, achieving state-of-the-art accuracy among privacy methods.

  15. LoMa: Local Feature Matching Revisited

    cs.CV 2026-04 unverdicted novelty 6.0

    Scaling data, model size, and compute for local feature matching produces large performance gains on challenging benchmarks and a new manually annotated HardMatch dataset.

  16. LoMa: Local Feature Matching Revisited

    cs.CV 2026-04 accept novelty 6.0

    Scaling data diversity, model capacity and compute for DeDoDe+LightGlue-style sparse matchers yields large gains and a new hard benchmark HardMatch.

  17. TerraSky3D: Multi-View Reconstructions of European Landmarks in 4K

    cs.CV 2026-03 unverdicted novelty 6.0

    TerraSky3D is a new high-resolution multi-view dataset with 50,000 images in 150 scenes of European landmarks, supplied with poses and depth maps to support 3D reconstruction research.

  18. AeroMap3D: Anchoring Monocular UAV 6-DoF Localization to Visual-Geometric-Semantic Map Priors

    cs.RO 2026-07 conditional novelty 5.0

    AeroMap3D combines a scale/yaw adapter, a frozen dense matcher, OSM-filtered DEM lifting, and an EKF to achieve 5.88 m mean 3D localization error over 55 km of UAV flight using only public maps.

  19. FreeTimeGS++: Secrets of Dynamic Gaussian Splatting and Their Principles

    cs.CV 2026-05 unverdicted novelty 5.0

    FreeTimeGS++ improves dynamic scene reconstruction by identifying emergent temporal partitioning and photometric-motion decoupling in 4DGS, then applying targeted techniques for reduced run-to-run variance.

  20. FreeTimeGS++: Secrets of Dynamic Gaussian Splatting and Their Principles

    cs.CV 2026-05 unverdicted novelty 5.0

    Analysis of 4DGS reveals temporal partitioning from Gaussian durations and a photometric-spatiotemporal discrepancy, leading to FreeTimeGS++ with gated marginalization and neural velocity fields for superior stability...

  21. FreeTimeGS++: Secrets of Dynamic Gaussian Splatting and Their Principles

    cs.CV 2026-05 unverdicted novelty 5.0

    FreeTimeGS++ improves 4D Gaussian Splatting by using gated marginalization and neural velocity fields to achieve more stable dynamic scene representations with lower run-to-run variance.

  22. Who Handles Orientation? Investigating Invariance in Feature Matching

    cs.CV 2026-04 accept novelty 5.0

    Learning rotation invariance in descriptors matches the performance of matcher-level invariance but allows earlier invariance, faster matchers, and no loss in upright performance when trained at scale.

Reference graph

Works this paper leans on

56 extracted references · 5 linked inside Pith · cited by 17 Pith papers

  1. [1]

    Map-free visual relocalization: Metric pose relative to a single image

    Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Dani- yar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. InEu- ropean Conf. Computer Vision (ECCV), 2022. 6

    2022

  2. [2]

    A general and adaptive robust loss func- tion

    Jonathan T Barron. A general and adaptive robust loss func- tion. InIEEE Conf. Computer Vision and Pattern Recogni- tion (CVPR), 2019. 5

    2019

  3. [3]

    Earthmatch: Iterative coregistration for fine-grained localization of astro- naut photography

    Gabriele Berton, Gabriele Goletto, Gabriele Trivigno, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthmatch: Iterative coregistration for fine-grained localization of astro- naut photography. InIEEE Conf. Computer Vision and Pat- tern Recognition (CVPR), 2024. 7, 2

    2024

  4. [4]

    Earthloc: Astronaut photography localization by indexing earth from space

    Gabriele Berton, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthloc: Astronaut photography localization by indexing earth from space. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  5. [5]

    Steerers: A framework for rotation equivariant keypoint descriptors

    Georg B ¨okman, Johan Edstedt, Michael Felsberg, and Fredrik Kahl. Steerers: A framework for rotation equivariant keypoint descriptors. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2024. 7

    2024

  6. [6]

    Vir- tual kitti 2.arXiv preprint arXiv:2001.10773, 2020

    Yohann Cabon, Naila Murray, and Martin Humenberger. Vir- tual kitti 2.arXiv preprint arXiv:2001.10773, 2020. 6

    Pith/arXiv arXiv 2001

  7. [7]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2020. 6

    2020

  8. [8]

    Brooks, Anton Van Den Hengel, and Darren Gawley

    Wojciech Chojnacki, Michael J. Brooks, Anton Van Den Hengel, and Darren Gawley. On the fitting of surfaces to data with covariances.IEEE Trans. Pattern Analysis and Machine Intelligence (T-PAMI), 22(11):1294–1303, 2000. 3

    2000

  9. [9]

    Scan- net: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scan- net: Richly-annotated 3d reconstructions of indoor scenes. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2017. 1, 7

    2017

  10. [10]

    Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization

    Siyan Dong, Shuzhe Wang, Shaohui Liu, Lulu Cai, Qingnan Fan, Juho Kannala, and Yanchao Yang. Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2025. 7

    2025

  11. [11]

    DKM: Dense kernelized feature matching for geometry estimation

    Johan Edstedt, Ioannis Athanasiadis, M ˚arten Wadenb ¨ack, and Michael Felsberg. DKM: Dense kernelized feature matching for geometry estimation. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 7

    2023

  12. [12]

    RoMa: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg B ¨okman, M ˚arten Wadenb¨ack, and Michael Felsberg. RoMa: Robust dense feature matching. InIEEE Conf. Computer Vision and Pat- tern Recognition (CVPR), 2024. 1, 3, 5, 6, 7

    2024

  13. [13]

    Virtual worlds as proxy for multi-object tracking analy- sis

    Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analy- sis. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016. 6

    2016

  14. [14]

    Cambridge university press,

    Richard Hartley and Andrew Zisserman.Multiple view ge- ometry in computer vision. Cambridge university press,

  15. [15]

    Image matching chal- lenge 2022, 2022

    Addison Howard, Eduard Trulls, Kwang Moo Yi, Dmitry Mishkin, Sohier Dane, and Yuhe Jin. Image matching chal- lenge 2022, 2022. 1

    2022

  16. [16]

    Uncertainty es- timates and multi-hypotheses networks for optical flow

    Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty es- timates and multi-hypotheses networks for optical flow. In European Conf. Computer Vision (ECCV), 2018. 5

    2018

  17. [17]

    Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018. 5

    Pith/arXiv arXiv 2018

  18. [18]

    MapAnything: Universal feed- forward metric 3D reconstruction, 2025

    Nikhil Keetha, Norman M ¨uller, Johannes Sch ¨onberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, Jonathon Luiten, Manuel Lopez-Antequera, Samuel Rota Bul`o, Christian Richardt, Deva Ramanan, Sebastian Scherer, and Peter Kontschieder. MapAnything: Universal feed- forward metric 3D reconstructio...

    Pith/arXiv arXiv 2025

  19. [19]

    Edgs: Eliminating densification for efficient convergence of 3dgs.arXiv preprint arXiv:2504.13204, 2025

    Dmytro Kotovenko, Olga Grebenkova, and Bj ¨orn Ommer. Edgs: Eliminating densification for efficient convergence of 3dgs.arXiv preprint arXiv:2504.13204, 2025. 1

    arXiv 2025

  20. [20]

    Dense-sfm: Structure from motion with dense consistent matching

    JongMin Lee and Sungjoo Yoo. Dense-sfm: Structure from motion with dense consistent matching. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2025. 1

    2025

  21. [21]

    Ground- ing image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing image matching in 3d with mast3r. InEuropean Conf. Computer Vision (ECCV), 2024. 3, 7

    2024

  22. [22]

    Megadepth: Learning single- view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single- view depth prediction from internet photos. InIEEE Conf. 9 Computer Vision and Pattern Recognition (CVPR), 2018. 1, 2, 6, 7

    2018

  23. [23]

    LightGlue: Local Feature Matching at Light Speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Polle- feys. LightGlue: Local Feature Matching at Light Speed. In IEEE Int’l Conf. Computer Vision (ICCV), 2023. 2, 7

    2023

  24. [24]

    Rubik: A struc- tured benchmark for image matching across geometric chal- lenges

    Thibaut Loiseau and Guillaume Bourmaud. Rubik: A struc- tured benchmark for image matching across geometric chal- lenges. InIEEE Conf. Computer Vision and Pattern Recog- nition (CVPR), 2025. 2

    2025

  25. [25]

    Opensfm.https : / / github

    Mapillary. Opensfm.https : / / github . com / mapillary/OpenSfM, 2014. 1

    2014

  26. [26]

    A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation

    Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2016. 6

    2016

  27. [27]

    WxBS: Wide baseline stereo generalizations

    Dmytro Mishkin, Jiri Matas, Michal Perdoch, and Karel Lenc. WxBS: Wide baseline stereo generalizations. In British Machine Vision Conference (BMVC), 2015. 1, 7, 8

    2015

  28. [28]

    Maxime Oquab, Timoth ´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Rus- sell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang- Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nico- las Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patri...

  29. [29]

    A guide to structureless visual localization

    V ojtech Panek, Qunjie Zhou, Yaqing Ding, S ´ergio Agostinho, Zuzana Kukelova, Torsten Sattler, and Laura Leal-Taix´e. A guide to structureless visual localization. arXiv preprint arXiv:2504.17636, 2025. 1

    Pith/arXiv arXiv 2025

  30. [30]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2021. 4, 1

    2021

  31. [31]

    Carl Edward Rasmussen and Christopher K. I. Williams. Gaussian Processes for Machine Learning (Adaptive Com- putation and Machine Learning). The MIT Press, 2005. 3

    2005

  32. [32]

    Susskind

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic syn- thetic dataset for holistic indoor scene understanding. In IEEE Int’l Conf. Computer Vision (ICCV), 2021. 4, 6, 8, 2

    2021

  33. [33]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7

    2020

  34. [34]

    Benchmarking 6dof outdoor visual localiza- tion in changing conditions

    Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6dof outdoor visual localiza- tion in changing conditions. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2018. 1

    2018

  35. [35]

    A benchmark and a baseline for robust multi- view depth estimation

    Philipp Schr ¨oppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multi- view depth estimation. InInt’l Conf. 3D Vision (3DV), 2022. 5

    2022

  36. [36]

    Structure-from-Motion Revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-Motion Revisited. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3

    2016

  37. [37]

    Oriane Sim ´eoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth´ee Darcet, Th´eo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie,...

    Pith/arXiv arXiv 2025

  38. [38]

    Mod- eling the world from internet photo collections.Int’l J

    Noah Snavely, Steven M Seitz, and Richard Szeliski. Mod- eling the world from internet photo collections.Int’l J. Com- puter Vision (IJCV), 80(2), 2008. 1, 3

    2008

  39. [39]

    Find my astronaut photo: Automated localization and georectification of astronaut photography

    Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR) Workshops, 2023. 7, 2

    2023

  40. [40]

    Roformer: Enhanced transformer with rotary position embedding, 2023

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2023. 6

    2023

  41. [41]

    LoFTR: Detector-free local feature match- ing with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature match- ing with transformers. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 7

    2021

  42. [42]

    Inloc: Indoor visual localization with dense matching and view synthesis

    Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak- ihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. InIEEE Conf. Computer Vi- sion and Pattern Recognition (CVPR), 2018. 1

    2018

  43. [43]

    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. 6

    2021

  44. [44]

    PDC-Net+: Enhanced Probabilistic Dense Cor- respondence Network.IEEE Trans

    Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. PDC-Net+: Enhanced Probabilistic Dense Cor- respondence Network.IEEE Trans. Pattern Analysis and Machine Intelligence (T-PAMI), 2023. 5

    2023

  45. [45]

    Aerialmegadepth: Learning aerial-ground reconstruction and view synthesis

    Khiem Vuong, Anurag Ghosh, Deva Ramanan, Srinivasa Narasimhan, and Shubham Tulsiani. Aerialmegadepth: Learning aerial-ground reconstruction and view synthesis. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2025. 6

    2025

  46. [46]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InIEEE Conf. Com- puter Vision and Pattern Recognition (CVPR), 2025. 3, 7, 5

    2025

  47. [47]

    Dust3r: Geometric 3d vi- sion made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vi- sion made easy. InIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3 10

    2024

  48. [48]

    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 Se- bastian Scherer. Tartanair: A dataset to push the limits of visual slam. InIEEE/RSJ Int’l Conf. Intelligent Robots and Systems (IROS), 2020. 6

    2020

  49. [49]

    Blendedmvs: A large-scale dataset for generalized multi-view stereo net- works

    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 net- works. InIEEE Conf. Computer Vision and Pattern Recog- nition (CVPR), 2020. 6

    2020

  50. [50]

    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. InIEEE Int’l Conf. Computer Vision (ICCV),

  51. [51]

    Hierarchical discrete distribution decomposition for match density esti- mation

    Zhichao Yin, Trevor Darrell, and Fisher Yu. Hierarchical discrete distribution decomposition for match density esti- mation. InIEEE Conf. Computer Vision and Pattern Recog- nition (CVPR), 2019. 5

    2019

  52. [52]

    Ufm: A simple path towards unified dense correspondence with flow

    Yuchen Zhang, Nikhil Keetha, Chenwei Lyu, Bhuvan Jhamb, Yutian Chen, Yuheng Qiu, Jay Karhade, Shreyas Jha, Yaoyu Hu, Deva Ramanan, Sebastian Scherer, and Wenshan Wang. Ufm: A simple path towards unified dense correspondence with flow. InAdvances in Neural Information Processing Systems (NeurIPS), 2025. 2, 3, 5, 6, 7

    2025

  53. [53]

    Efficient motion prompt learning for robust visual tracking

    Jie Zhao, Xin Chen, Yongsheng Yuan, Michael Felsberg, Dong Wang, and Huchuan Lu. Efficient motion prompt learning for robust visual tracking. InInt’l Conf. Machine learning (ICML), 2025. 6

    2025

  54. [54]

    Xiaoming Zhao, Xingming Wu, Weihai Chen, Peter C. Y . Chen, Qingsong Xu, and Zhengguo Li. Aliked: A lighter keypoint and descriptor extraction network via deformable transformation.IEEE Transactions on Instrumentation & Measurement, 72, 2023. 5

    2023

  55. [55]

    match embeddings

    Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. InEuropean Conf. Computer Vision (ECCV), 2018. 5 11 RoMa v2: Harder Better Faster Denser Feature Matching Supplementary Material A. Architectural Details Here we give further details on the exact dimensions of in- puts and outputs of the different components of our mo...

    2018

  56. [56]

    TRAIN” “TEST

    This has the effect of making the refinement significantly cheaper, as we also only have to extract features from the VGG19 backbone until stride 4, compared to RoMa and DKM which require features and refinement from stride16. We denote the fine features as φ4 ∈R H 4 × W 4 ×192, φ2 ∈R H 2 × W 2 ×48, φ1 ∈R H×W×12 , (15) where the dimensions come from the r...