Augmented Reality in Cultural Heritage: A Dual-Model Pipeline for 3D Artwork Reconstruction

Alessia Castronovo; Claudio Piciarelli; Daniele Pannone; Danilo Avola; Gian Luca Foresti; Maurizio Mancini; Muhammad Yasir Bilal; Rossana Gabrieli

REVIEW 4 major objections 6 minor 1 cited by

Augmented Reality in Cultural Heritage: A Dual-Model Pipeline for 3D Artwork Reconstruction

T0 review · 4 major / 6 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 The paper claims that fusing GLPN and Depth-Anything depth maps with a 0.97/0.03 weighted average, followed by range normalization, yields single-image 3D artwork reconstructions whose CLIP similarity to the original exceeds either model…

desk verdict A clearly written museum-AR pipeline that combines two off-the-shelf depth models, but the reported accuracy gain rests on a CLIP-similarity proxy over five tuned images, so the central claim is not yet established. read the letter →

arxiv 2507.13719 v1 pith:YKAKOJKL submitted 2025-07-18 cs.CV

Daniele Pannone , Alessia Castronovo , Maurizio Mancini , Gian Luca Foresti , Claudio Piciarelli , Rossana Gabrieli , Muhammad Yasir Bilal , Danilo Avola This is my paper · ORCID

classification cs.CV

keywords augmentedrealityculturalheritagemonoculardepthestimationsingle-image3DreconstructionGLPNDepth-AnythingCLIPsimilarityPoissonsurface

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 sets out to show that a single photograph of an artwork can be turned into a 3D mesh good enough for augmented reality in a museum, without photogrammetry or depth sensors. Its route is to fuse two off-the-shelf monocular depth models: GLPN, which keeps global depth layers, and Depth-Anything, which supplies fine local detail, combined in a $0.97/0.03$ weighted average and then range-normalized. The fused depth map is back-projected into a point cloud, cleaned, and meshed with Poisson surface reconstruction before being displayed in Unity. The paper's evidence is CLIP semantic-visual similarity between each original painting and a lateral render of the corresponding mesh, where the fusion outscores either model alone. A reader should care because, if the claim holds, existing flat images of artworks could power interactive 3D museum experiences at low capture cost.

What carries the argument

The load-bearing mechanism is the fused, normalized depth map. GLPN's output $D_{\text{GLPN}}$ and Depth-Anything's resized output $\tilde{D}_{\text{DA}}$ are combined pixel-wise as $D_{\text{combined}} = \alpha D_{\text{GLPN}} + (1-\alpha)\tilde{D}_{\text{DA}}$ with $\alpha=0.97$, then min-max scaled into $[0.6,1.0]$ to cancel the models' different depth calibrations. This map is back-projected through the pinhole camera model into a point cloud, cleaned by statistical outlier removal, supplied with estimated normals, and converted into a watertight mesh by solving the Poisson equation $\Delta\chi = \nabla \cdot \mathbf{V}$ for the surface indicator function. The mesh is manually segmented in Blender and loaded into Unity, where the AR Tracked Image Manager anchors it to the recognized artwork image.

What would settle it

Render each generated mesh from the same camera viewpoint as the original artwork, obtain a geometric ground truth for the same paintings (from photogrammetry, structure from motion, or a synthetic painting-like scene with known depth), and compare the fused pipeline's depth error or point-cloud distance against the two single models. If the fused depth maps do not have lower geometric error than GLPN or Depth-Anything alone, the paper's claim of improved reconstruction accuracy is falsified even if CLIP similarity still rises.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that two pre-trained monocular depth estimators fail in complementary ways, and that a nearly all-GLPN fusion corrects both. GLPN alone flattens local details such as facial features; Depth-Anything alone produces disproportionate shapes because its depth scale is inconsistent. The proposed convex combination $D_{\text{combined}} = \alpha D_{\text{GLPN}} + (1-\alpha)\tilde{D}_{\text{DA}}$ with $\alpha = 0.97$, followed by normalization to $[0.6,1.0]$, keeps GLPN's global structure while importing enough Depth-Anything detail to improve perceptual fidelity. Table 1 reports CLIP similarity scores for five test works -- Lady with an Ermine, The Starry Night, Mona Lisa, Relativity, and Girl with a Pearl Earring -- and in every row the proposed pipeline scores higher than either single model, with Mona Lisa rising to $0.7686$ from $0.5732$ (Depth-Anything) and $0.5891$ (GLPN). The paper concludes that the fused pipeline is semantically and visually more consistent with the original artwork when the reconstruction is viewed from a partially lateral angle.

Load-bearing premise

The load-bearing premise is that the visual-similarity score produced by the CLIP image encoder, comparing the original painting to a rendered side view of the mesh, is a valid measure of 3D reconstruction quality; the paper adopts this proxy because direct geometric measurements were not available.

Editorial extensions

If this is right

A museum visitor could point a phone at a painting and see a 3D model of it anchored in space, generated on the fly from that single camera frame.
Institutions with only flat archival photographs could generate interactive 3D content without multi-view capture rigs or depth sensors.
Because only three percent of Depth-Anything signal is needed to improve perceived fidelity, the fusion leverages detail-rich models without sacrificing global coherence.
The CLIP-based protocol gives a quantitative way to compare reconstruction variants when no geometric ground truth exists, and it favors the fused pipeline consistently across five artworks.
The method is most reliable for portraits, where painted shading and contrast supply depth cues, and least reliable for soft-brushstroke landscapes and optical illusions.

Reading between the lines

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

The reported 'reconstruction accuracy' is measured by CLIP similarity, not geometry; a lateral render of a badly distorted mesh could still match the original painting in CLIP space, so the actual 3D metric accuracy of the fusion remains untested.
A direct geometric test would be the natural extension: run the same fusion alongside a structure-from-motion or photogrammetric reconstruction of the same artworks and compare chamfer distance or depth error; only then would the fusion's geometric superiority be established.
The single fixed weight $\alpha=0.97$ was chosen on a five-image set; on abstract or low-texture artworks, where Depth-Anything's detail signal is sparse, the optimal balance could shift, and an adaptive weight per image would be a plausible improvement.
The pipeline could be extended to fuse depth estimates across multiple phone viewpoints over time, keeping single-image simplicity while incrementally approaching the accuracy of multi-view methods.

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

What you should know up front: this paper is a straightforward engineering pipeline. It takes GLPN and Depth-Anything depth maps, averages them with alpha=0.97, normalizes to [0.6, 1.0], backs the depth into a point cloud, runs Poisson meshing, and drops the mesh into Unity AR. That is the whole novel content. It is a new application and a routine extension of existing components. The paper also describes the steps clearly and does not hide its limitations — it explicitly says direct geometric evaluation was not feasible, and it acknowledges that single-image reconstruction lacks depth cues for landscapes. No equations are wrong, and the writing is honest.

The soft spot is the evaluation. Table 1 reports CLIP cosine similarity between the original artwork and a lateral render of the mesh. CLIP is trained for image-text alignment, not 3D structure, so a textured lateral view of a nearly flat mesh can score high even if the geometry is wrong. The table has five rows, no variance, no significance test, and alpha was chosen while testing on those very five images. The paper also does not say whether the GLPN and Depth-Anything baselines received the same range normalization and Blender post-processing as the proposed pipeline. Those confounds make the claimed 'significant improvements' hard to take seriously. The stress-test note about the baselines is fair: without controlling for those steps, the comparison is not apples-to-apples.

That said, this is not a reject-on-sight paper. The pipeline is plausible, the domain is legitimate, and the limitations are stated rather than hidden. The problem is that the evidence does not support the abstract's claim about reconstruction accuracy. It supports only that a CLIP model, on five famous paintings, rated one textured render more similar than two others. That is a much weaker statement.

If you are a referee, you should send this back for major revision, not throw it out. The authors need to either obtain real 3D ground truth (even a few scans of bas-reliefs or sculptures), or clearly argue why CLIP is a valid proxy, and they need to evaluate on more than five images with statistics and equal treatment of baselines. Without that, the contribution is just a demo. For a workshop or an applied cultural-heritage venue, it might be acceptable; for a mainline computer-vision conference, the evidence bar is too low.

My bottom line: I would accept the paper for peer review because the writing is clear and the problem is real, but I would expect heavy revision and a much more careful evaluation before publication.

Referee Report

4 major / 6 minor

Summary. The paper proposes a pipeline for generating 3D meshes of artworks from a single image: two pretrained monocular depth models (GLPN and Depth-Anything) are fused by a weighted average (Eq. 2) with alpha = 0.97, the resulting depth is range-normalized (Eq. 3), back-projected into a point cloud, converted to a mesh via Poisson surface reconstruction, and finally visualized in an AR application built with Unity. The quantitative evaluation in Sec. 4.2 uses CLIP cosine similarity between the original artwork image and a lateral render of the reconstructed mesh, reported for five artworks in Table 1, with qualitative comparison in Fig. 2. The abstract claims that experimental results demonstrate significant improvements in reconstruction accuracy and visual realism from the dual-model fusion.

Significance. If the reported gains were established, the work would offer a practical single-image AR reconstruction pipeline for museum settings, building on readily available pretrained models and providing a full integration into an AR framework. The paper is transparent about a key limitation, stating in Sec. 4.2 that direct geometric evaluation was not feasible, and it describes the implementation in enough detail to be reimplemented. However, the quantitative evidence is a non-geometric CLIP similarity metric, and the fusion weight is selected on the same five test artworks, so the central claim of improved reconstruction accuracy is not currently supported. The significance therefore rests on whether the authors can replace or supplement the CLIP-based evaluation with a geometric or at least statistically robust validation.

major comments (4)

[Sec. 4.2, Table 1] The abstract's claim of "significant improvements in reconstruction accuracy" is not supported by the reported evidence, which measures only CLIP cosine similarity between the original artwork and a lateral render of the mesh (Eqs. 7-9). The paper itself states in Sec. 4.2 that "direct geometric evaluation was not feasible." CLIP embeddings are trained for image-text alignment and are largely invariant to 3D structure, so a textured but geometrically implausible mesh can score highly. A geometric metric (e.g., depth error on synthetic ground-truth data, multi-view consistency, or a clear comparison against a planar baseline) or alternatively a reframed claim about "perceptual similarity" is needed to support the stated conclusion.
[Sec. 3.2, Eq. (2)] The fusion weight alpha = 0.97 is "empirically determined" during testing, and the same five artworks then appear in Table 1 as the evidence for improvement. This makes the reported gain partly a fitted result. The authors should either fix alpha on a validation set disjoint from the test artworks or report a sensitivity analysis over alpha values and a held-out set of artworks.
[Sec. 3.2-3.3 vs. Sec. 4.2] The comparison in Table 1 may be unfair because the proposed pipeline includes not only the fusion but also range normalization (Eq. 3), back-projection with assumed camera intrinsics, statistical outlier removal, Poisson reconstruction, and manual Blender segmentation and material assignment described in Sec. 3.2. The manuscript does not state whether these exact post-processing steps were applied to the Depth-Anything-only and GLPN-only baselines. If they were not, the improvement may reflect unequal preprocessing rather than the fusion itself. Identical post-processing for all three variants must be applied and reported.
[Table 1] The quantitative evaluation has no measure of variability: five artworks, one rendered view per model, no error bars, and no statistical test. CLIP similarity can fluctuate with viewpoint and rendering settings, and the claimed gains (e.g., Mona Lisa 0.5891 to 0.7686) could arise from a single favorable view. Multiple renders per mesh (varying lateral angle and lighting) and a paired statistical test would be needed to support the statement that the proposed pipeline "consistently yields higher similarity scores."

minor comments (6)

[Sec. 4.1] There is a typo "the the used tool" in the first sentence of Sec. 4.1, and the text refers to "T able 1" before the table appears.
[Eqs. (4) and (5)] Equations (4) and (5) express the same back-projection operation with different notations (K^-1 times pixel coordinates vs. the pinhole formulas). The paper should use one consistent notation and clearly define the relationship between K and (fx, fy, cx, cy).
[Sec. 3.2] The notation for the depth maps is inconsistent: the text writes "DGLP N" with an extra space, and the dimensions H' and W' for Depth-Anything are introduced but not used consistently in Eq. (2).
[Sec. 4.2] Equation (9) defines a mean similarity score, but Table 1 reports only individual artwork scores and does not include the mean; reporting the mean with standard deviation would aid interpretation.
[Fig. 2] The qualitative comparison is difficult to assess from a single 2D figure; a supplementary video or interactive 3D model viewer would allow reviewers to verify the claimed visual improvements.
[Sec. 4.2] The rendering conditions for the mesh views (camera pose, lighting, background, and exact lateral angle) are not specified, so the CLIP scores in Table 1 are not reproducible as reported.

Circularity Check

1 steps flagged · score 6.0 of 10

Fusion weight α is tuned during testing and then evaluated on the same artworks, making the reported accuracy gain partly a fitted in-sample result.

fitted input called prediction [Section 3.2, Eq. (2); Section 4.2, Table 1]
"where α = 0.97 was empirically determined to preserve the global structural consistency of GLPN while incorporating local detail enhancements from Depth-Anything. Notably, even a small contribution from Depth-Anything 1 − α = 0.03 was found to significantly improve the perceptual fidelity of the final depth map."

The convex weight α in Eq. (2) is stated to be 'empirically determined' during testing, with no reported held-out validation split. Table 1 then reports that the fused pipeline (α=0.97) outperforms the two endpoint models (α=1 for GLPN and α=0 for Depth-Anything) on exactly the artworks used for the evaluation. Because Eq. (2) is a convex combination, choosing α on the test set guarantees that the reported optimum is at least as good as either endpoint on that set; the claimed 'significant improvements in reconstruction accuracy' are therefore an in-sample consequence of the tuning decision rather than an out-of-sample prediction. The paper does not establish that α was fixed before seeing the test artworks or that the CLIP comparison was performed on independent data.

full rationale

The pipeline's external components are independent: GLPN and Depth-Anything are pre-trained models, and the back-projection (Eqs. 4–5), Poisson reconstruction (Eq. 6), and AR assembly are standard operations. No load-bearing self-citation chain or imported uniqueness theorem appears. The main circularity is the tuning of α=0.97 in Section 3.2 on the test corpus, followed by Table 1 reporting improvements on the same corpus; since α is a convex weight, the reported gain over the pure GLPN and Depth-Anything baselines is partly forced by the in-sample choice. Additionally, the quantitative evaluation uses CLIP cosine similarity (Eqs. 7–9) between the original artwork and a rendered lateral view, not a geometric error; the paper itself admits 'direct geometric evaluation was not feasible.' This weakens the abstract's 'reconstruction accuracy' claim, though it is more a metric-validity limitation than a circular derivation. The baseline comparison may also be confounded by the proposed pipeline's extra normalization (Eq. 3) and manual Blender segmentation, but that is a fairness issue rather than circularity. Overall, the central quantitative claim reduces in part to a test-set-tuned parameter, giving partial circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The quantitative claims rest on two external pretrained models and the CLIP proxy rather than on new theory. The only fitted component is the fusion weight, supplemented by hand-chosen normalization and assumed camera intrinsics; no new entities are introduced.

free parameters (4)

Fusion weight alpha = 0.97
Chosen empirically during testing (Section 3.2, Eq 2); the same five artworks are later used to report the improvement in Table 1, so the gain is partly fitted.
Depth normalization target interval [d_min, d_max] = [0.6, 1.0]
Hand-selected in Eq 3 to reconcile incompatible depth scales; no sensitivity analysis or independent validation is reported.
Camera intrinsic parameters (fx, fy, cx, cy) = not specified
Eqs 4 and 5 back-project using a pinhole camera matrix K, but the paper does not state how K is chosen for arbitrary artwork images; different intrinsics change the geometry of the point cloud.
Outlier removal and Poisson reconstruction settings = not reported
Section 3.2 invokes statistical outlier removal and Section 3.3 Poisson surface reconstruction without thresholds or octree depth; these affect smoothness, detail, and mesh closure.

assumptions (5)

domain assumption Pretrained depth models transfer to artworks
GLPN is trained on NYU Depth V2 and Depth-Anything on NYUv2/KITTI (Section 3.1); paintings contain irregular contours and brushstroke textures outside this distribution, and Section 4.3 admits landscape paintings are harder.
domain assumption CLIP similarity is a valid proxy for 3D reconstruction quality
Section 4.2 adopts CLIP because 'direct geometric evaluation was not feasible'; CLIP is an image-text embedding model, not a geometric accuracy measure, so this assumption carries the quantitative claim.
domain assumption Normalized relative depths can be treated as Euclidean distances
Eqs 2-4 rescale relative depth into [0.6, 1.0] and back-project with a pinhole model without validating that the normalized values correspond to physical distances.
standard math Poisson reconstruction yields faithful surfaces for artistic subjects
Section 3.3 uses the standard Poisson formulation (Eq 6); the math is standard, but it assumes oriented points and a watertight surface, which may be violated for thin or textured paintings.
domain assumption Five selected artworks are representative of museum collections
Section 4.1 tests five manually chosen paintings; there is no diversity sampling, no failure distribution, and no statistical power.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Augmented Reality in Cultural Heritage: A Dual-Model Pipeline for 3D Artwork Reconstruction." pith.science (2026). https://pith.science/paper/YKAKOJKL

@misc{pith2026250713719,
  author       = {Pith},
  title        = {Pith review of: Augmented Reality in Cultural Heritage: A Dual-Model Pipeline for 3D Artwork Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKAKOJKL}},
  note         = {Machine review of arXiv:2507.13719}
}

read the original abstract

This paper presents an innovative augmented reality pipeline tailored for museum environments, aimed at recognizing artworks and generating accurate 3D models from single images. By integrating two complementary pre-trained depth estimation models, i.e., GLPN for capturing global scene structure and Depth-Anything for detailed local reconstruction, the proposed approach produces optimized depth maps that effectively represent complex artistic features. These maps are then converted into high-quality point clouds and meshes, enabling the creation of immersive AR experiences. The methodology leverages state-of-the-art neural network architectures and advanced computer vision techniques to overcome challenges posed by irregular contours and variable textures in artworks. Experimental results demonstrate significant improvements in reconstruction accuracy and visual realism, making the system a highly robust tool for museums seeking to enhance visitor engagement through interactive digital content.

Figures

Figures reproduced from arXiv: 2507.13719 by the authors.

**Figure 1.** Overview of the proposed pipeline: (a) The input image is processed to generate a depth map using the GLPN and Depth-Anything models. The resulting depth map is used to compute a point cloud (b), which is then converted into a 3D mesh (c) and finally visualized in augmented reality (d). GLPN and Depth-Anything, respectively. As the outputs may differ in spatial resolution, DDA was first resized to the resolution of … view at source ↗

**Figure 2.** Qualitative comparison of the 3D mesh reconstruction results, produced using Depth-Anything, GLPN and the proposed pipeline respectively. the subject depicted. Although some variation in accuracy is observed depending on the pictorial detail, the pipeline proved capable of adapting effectively to a range of artistic styles. Moreover, the proposed solution outperformed the individual use of the two depth estimation … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Toward Site-Aware MR Art Exhibitions: A SLAM-Based Deployment Pipeline for Spatial Coherence and Exhibition Experience
cs.MM 2026-07 conditional novelty 5.0 of 10

A SLAM-based pipeline for large-scale MR art exhibitions was deployed and positively rated, but its main experiential claims rest on a small pilot and a baseline-free user study.

Reference graph

Works this paper leans on

21 extracted references · 14 canonical work pages · cited by 1 Pith paper

[1]

In: Sclaroff, S., Distante, C., Leo, M., Farinella, G.M., Tombari, F

Avola, D., Cinque, L., Fagioli, A., Foresti, G.L., Marini, M.R., Mecca, A., Pannone, D.: Medicinal boxes recognition on a deep transfer learning augmented reality mobile application. In: Sclaroff, S., Distante, C., Leo, M., Farinella, G.M., Tombari, F. (eds.) Image Analysis and Processing – ICIAP 2022. pp. 489–499. Springer International Publishing, Cham (2022)

work page 2022
[2]

In: Ricci, E., Rota Bulò, S., Snoek, C., Lanz, O., Messelodi, S., Sebe, N

Avola, D., Cinque, L., Fagioli, A., Foresti, G.L., Massaroni, C., Pannone, D.: Feature-based slam algorithm for small scale uav with nadir view. In: Ricci, E., Rota Bulò, S., Snoek, C., Lanz, O., Messelodi, S., Sebe, N. (eds.) Image Analy- sis and Processing – ICIAP 2019. pp. 457–467. Springer International Publishing, Cham (2019)

work page 2019
[3]

Sensors 20(18) (2020)

Avola, D., Cinque, L., Fagioli, A., Foresti, G.L., Pannone, D., Piciarelli, C.: Bodyprint—a meta-feature based lstm hashing model for person re-identification. Sensors 20(18) (2020)

work page 2020
[4]

In: De Marsico, M., di Baja, G.S., Fred, A

Avola,D.,Cinque,L.,Foresti,G.L.,Pannone,D.:Visualcryptographyfordetecting hidden targets by small-scale robots. In: De Marsico, M., di Baja, G.S., Fred, A. (eds.) Pattern Recognition Applications and Methods. pp. 186–201. Springer International Publishing, Cham (2019)

work page 2019
[5]

language-image learning

Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive 12 Pannone et al. language-image learning. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 2818–2829 (2023)

work page 2023
[6]

IEEE Transactions on Intelligent Transportation Systems26(3), 2967–2976 (2025)

Corno, M., Franceschetti, L., Matteo Savaresi, S.: Design of a cost effective spatial image registration system for augmented reality in vehicular applications. IEEE Transactions on Intelligent Transportation Systems26(3), 2967–2976 (2025)

work page 2025
[7]

International Journal of Robotics Research (IJRR) (2013)

Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR) (2013)

2013
[8]

Springer Cham, 1 edn

Ikeuchi, K., Matsushita, Y., Sagawa, R., Kawasaki, H., Mukaigawa, Y., Furukawa, R., Miyazaki, D.: Active Lighting and Its Application for Computer Vision. Springer Cham, 1 edn. (2020)

work page 2020

Show all 21 references

[9]

In: Sheffer, A., Polthier, K

Kazhdan, M., Bolitho, M., Hoppe, H.: Poisson surface reconstruction. In: Sheffer, A., Polthier, K. (eds.) Symposium on Geometry Processing. The Eurographics Association (2006)

2006
[10]

ACM Transactions on Graphics42(4) (2023)

Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics42(4) (2023)

2023
[11]

CoRRabs/2201.07436 (2022), https://arxiv.org/abs/2201.07436

Kim, D., Ka, W., Ahn, P., Joo, D., Chun, S., Kim, J.: Global-local path networks for monocular depth estimation with vertical cutdepth. CoRRabs/2201.07436 (2022), https://arxiv.org/abs/2201.07436

2022 arXiv
[12]

In: 2023 International Conference on Electronics, Information, and Communication (ICEIC)

Kim, K., Kim, M.J., Kim, H., Park, S., Paik, J.: Person re-identification method us- ing text description through clip. In: 2023 International Conference on Electronics, Information, and Communication (ICEIC). pp. 1–4 (2023)

2023
[13]

Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W.,Howes,R.,Huang,P.Y.,Li,S.W.,Misra,I.,Rabbat,M.,Sharma,V.,Synnaeve, G., Xu, H., Jegou, H., Mairal, J., Labatut, P....

2024 arXiv
[14]

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021),https://arxiv.org/abs/ 2103.00020

2021 arXiv
[15]

In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV)

Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12159–12168 (2021)

2021
[16]

The Pho- togrammetric Record 21(115), 269–291 (2006)

Remondino, F., El-Hakim, S.: Image-based 3d modelling: A review. The Pho- togrammetric Record 21(115), 269–291 (2006)

2006
[17]

In: Fitzgibbon, A., Lazebnik, S., Perona, P., Sato, Y., Schmid, C

Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from rgbd images. In: Fitzgibbon, A., Lazebnik, S., Perona, P., Sato, Y., Schmid, C. (eds.) Computer Vision – ECCV 2012. pp. 746–760. Springer Berlin Heidelberg, Berlin, Heidelberg (2012)

2012
[18]

IEEE Transactions on Image Processing 28(1), 394–409 (2019)

Tan, W.R., Chan, C.S., Aguirre, H., Tanaka, K.: Improved artgan for conditional synthesis of natural image and artwork. IEEE Transactions on Image Processing 28(1), 394–409 (2019)

2019
[19]

arXiv preprint arXiv:2403.02151 (2024)

Tochilkin, D., Pankratz, D., Liu, Z., Huang, Z., , Letts, A., Li, Y., Liang, D., Laforte, C., Jampani, V., Cao, Y.P.: Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151 (2024)

2024 arXiv
[20]

In: 2016 IEEE Conference on Computer Vision and Pattern Recog- nition Workshops (CVPRW)

Tron, R., Zhou, X., Daniilidis, K.: A survey on rotation optimization in structure from motion. In: 2016 IEEE Conference on Computer Vision and Pattern Recog- nition Workshops (CVPRW). pp. 1032–1040 (2016)

2016
[21]

In: CVPR (2024)

Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Un- leashing the power of large-scale unlabeled data. In: CVPR (2024)

2024

Pith tools

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

[1] [1]

In: Sclaroff, S., Distante, C., Leo, M., Farinella, G.M., Tombari, F

Avola, D., Cinque, L., Fagioli, A., Foresti, G.L., Marini, M.R., Mecca, A., Pannone, D.: Medicinal boxes recognition on a deep transfer learning augmented reality mobile application. In: Sclaroff, S., Distante, C., Leo, M., Farinella, G.M., Tombari, F. (eds.) Image Analysis and Processing – ICIAP 2022. pp. 489–499. Springer International Publishing, Cham (2022)

work page 2022

[2] [2]

In: Ricci, E., Rota Bulò, S., Snoek, C., Lanz, O., Messelodi, S., Sebe, N

Avola, D., Cinque, L., Fagioli, A., Foresti, G.L., Massaroni, C., Pannone, D.: Feature-based slam algorithm for small scale uav with nadir view. In: Ricci, E., Rota Bulò, S., Snoek, C., Lanz, O., Messelodi, S., Sebe, N. (eds.) Image Analy- sis and Processing – ICIAP 2019. pp. 457–467. Springer International Publishing, Cham (2019)

work page 2019

[3] [3]

Sensors 20(18) (2020)

Avola, D., Cinque, L., Fagioli, A., Foresti, G.L., Pannone, D., Piciarelli, C.: Bodyprint—a meta-feature based lstm hashing model for person re-identification. Sensors 20(18) (2020)

work page 2020

[4] [4]

In: De Marsico, M., di Baja, G.S., Fred, A

Avola,D.,Cinque,L.,Foresti,G.L.,Pannone,D.:Visualcryptographyfordetecting hidden targets by small-scale robots. In: De Marsico, M., di Baja, G.S., Fred, A. (eds.) Pattern Recognition Applications and Methods. pp. 186–201. Springer International Publishing, Cham (2019)

work page 2019

[5] [5]

language-image learning

Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive 12 Pannone et al. language-image learning. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 2818–2829 (2023)

work page 2023

[6] [6]

IEEE Transactions on Intelligent Transportation Systems26(3), 2967–2976 (2025)

Corno, M., Franceschetti, L., Matteo Savaresi, S.: Design of a cost effective spatial image registration system for augmented reality in vehicular applications. IEEE Transactions on Intelligent Transportation Systems26(3), 2967–2976 (2025)

work page 2025

[7] [7]

International Journal of Robotics Research (IJRR) (2013)

Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The kitti dataset. International Journal of Robotics Research (IJRR) (2013)

2013

[8] [8]

Springer Cham, 1 edn

Ikeuchi, K., Matsushita, Y., Sagawa, R., Kawasaki, H., Mukaigawa, Y., Furukawa, R., Miyazaki, D.: Active Lighting and Its Application for Computer Vision. Springer Cham, 1 edn. (2020)

work page 2020