Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Methods for evaluating the resolution of 3D data derived from satellite images

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read An adapted contrast transfer function measures the horizontal resolution of satellite-derived 3D data from parallel building pairs.

desk verdict Genuinely new CTF adaptation for 3D satellite-derived data, but the headline distances hinge on a Gaussian model validated only on synthetic downsampling—worth refereeing with requests for UQ and robustness checks. read the letter →

arxiv 2506.11876 v1 pith:XX2QGD7F submitted 2025-06-13 cs.CV eess.IV

Christina Selby , Holden Bindl , Tyler Feldman , Andrew Skow , Nicolas Norena Acosta , Shea Hagstrom , Myron Brown This is my paper · ORCID
classification cs.CVeess.IV
keywords contrasttransferfunction3Dresolutionsatelliteimagerydigitalsurfacemodelpointcloudbuildingfootprintreferencelidarmulti-viewstereo
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 seeks a way to measure the horizontal resolution of 3D products reconstructed from satellite imagery, including point clouds, digital surface models, and meshes, without requiring semantic labels. Its proposal is to adapt the contrast transfer function (CTF) used in imaging to elevation data: pairs of parallel buildings play the role of a tribar target, and the contrast between building tops and the ground between them is tracked as a function of building separation. The measured contrast is modeled by $C(d)=A\exp(-(\pi\sigma/d)^2)$, and the distance where that fitted curve crosses a chosen CTF threshold is reported as the resolution. On test DSMs derived from commercial satellite imagery, the method gives about 2.5 m over Nellis Air Force Base and about 2 m over Jacksonville. If the method is right, it gives analysts a way to compare reconstruction processes and track resolution improvements without custom in-scene targets.

What carries the argument

The central object is the adapted contrast transfer function for elevation, $C(d)=0.5\left(\frac{A_1-B}{A_1+B}+\frac{A_2-B}{A_2+B}\right)$, computed over evaluation regions formed by parallel building footprints. The fitting model $C(d)=A\exp(-(\pi\sigma/d)^2)$ is the mechanism that converts scattered CTF measurements into a single resolution distance: it assumes the effective point-spread function of the 3D reconstruction process is Gaussian, and the threshold crossing of the fitted curve is declared the distance at which buildings are resolved.

What would settle it

Measure the same test product's actual edge response at a known building edge, convert that edge-spread width to a contrast threshold distance, and compare it with the distance read from the fitted Gaussian curve; disagreement would mean the reported resolution inherits the model's shape.

Watch

Extended reading notes

Core claim

The central claim is that a contrast transfer function computed from elevation, rather than image intensity, quantifies the horizontal resolution of general 3D data products. The pipeline builds evaluation regions from building footprint pairs: a center rectangle over the ground and two adjacent rectangles over the buildings; after local alignment of test and reference elevations, the CTF value is the average of the contrasts from each building region against the ground region. Reference lidar supplies the DSM and building footprints, and the test product is aligned to it. The scatter of CTF versus building-pair distance is fit with $C(d)=A\exp(-(\pi\sigma/d)^2)$, with $A$ and $\sigma$ fitted, and the resolution is read where the curve meets a CTF threshold, typically 0.2. The model choice is validated on synthetic DSMs progressively downsampled so that the expected CTF=0.2 distance doubles with each step. At the two test sites this yields roughly 2.5 m at Nellis and roughly 2 m at Jacksonville, and the authors report that lidar-derived and hand-curated footprint sources agree closely.

Load-bearing premise

The load-bearing premise is that the fitted curve $C(d)=A\exp(-(\pi\sigma/d)^2)$ captures the true fall-off of elevation contrast with distance, because the reported resolution values are read from that fitted curve rather than measured independently.

Editorial extensions

If this is right

  • The same pipeline can compare different 3D reconstruction processes by running each on the same set of sites and comparing their threshold distances.
  • It can also compare one reconstruction process run with different input imagery, isolating the effect of the inputs on horizontal resolution.
  • Because the test and reference data are left aligned at the pixel level, the same workflow supports vertical accuracy statistics such as RMSE and percentile errors alongside the CTF resolution.
  • Tracking the fitted resolution over time gives a quantitative way to measure whether new 3D reconstruction algorithms are actually producing sharper results.
  • Areas with plentiful, unobstructed parallel buildings become reusable test sites, removing the need to deploy physical tribar targets.

Reading between the lines

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

  • A next step beyond the paper would be to validate the fitted $\sigma$ against an independent edge-spread measurement on the same product, rather than only against downsampled synthetic DSMs.
  • The reported distances inherit the Gaussian point-spread assumption; if a particular reconstruction has a different blur shape, the threshold distance could still be a fair relative comparator but would not be the literal physical resolution stated by the fitted curve.
  • The same contrast logic could be flipped to measure the minimum discernible object size, which the authors list as future work, by using single buildings as targets instead of building pairs.
  • Relating the fitted resolution to satellite viewing geometry and image metadata would allow operators to predict which parts of a scene a reconstruction will resolve sharply.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces an adaptation of the contrast transfer function (CTF) for evaluating the horizontal resolution of 3D data products (point clouds, DSMs, meshes) derived from satellite imagery. The method uses parallel building pairs as in-scene proxies for tribar targets: for each pair, it computes a contrast between building and ground elevations in the test data relative to a reference lidar-derived DSM, after local alignment and normalization. The resulting contrast values are plotted against building separation distance and fitted to the model C(d) = A exp(-(pi sigma/d)^2) (Eq. 2). The horizontal resolution is then reported as the distance at which this fitted curve crosses a chosen CTF threshold (typically 0.2). The pipeline includes data preparation, phase-correlation alignment, building footprint generation (from lidar segmentation or OSM), evaluation-region construction, and CTF calculation. Results are presented for two sites, Nellis AFB and Jacksonville, reporting threshold distances of approximately 2.5 m and 2 m, respectively, with comparisons between footprint sources and reference-CTF filters.

Significance. If it holds up, the method fills a genuine gap: existing resolution metrics for 3D data require semantic labels or manual polygons, whereas this approach is designed for general point clouds, DSMs, and meshes without such labels. The use of in-scene building pairs is a practical alternative to artificial calibration targets and could be useful for comparing reconstruction pipelines or input imagery. The paper gives a detailed, reproducible-sounding pipeline and includes a synthetic sanity check (Figure 1) showing that the model form captures downsampling behavior. The inclusion of two real sites and multiple footprint sources is a strength. However, the central resolution values are model outputs from an unverified Gaussian-PSF assumption, and no uncertainty quantification is provided, so the quantitative claims are not yet fully supported.

major comments (4)
  1. [Section 2.1, Eq. (2), Figure 1] The model C(d) = A exp(-(pi sigma/d)^2) is the load-bearing assumption of the method, yet it is validated only on synthetic DSMs progressively downsampled by factors of two. Downsampling is a specific isotropic low-pass operation that is not representative of the full error characteristics of MVS or NeRF reconstructions (e.g., matching artifacts, depth discontinuities, anisotropy). The paper does not provide evidence that real 3D reconstruction point-spread functions are approximately Gaussian. As a result, the reported threshold distances (about 2.5 m at Nellis and 2 m at Jacksonville) are conditional on an unverified model form. I recommend adding a validation experiment in which a real DSM is degraded by a controlled, known blur (e.g., Gaussian with a range of sigma values) and comparing the method's inferred threshold to the truth, and also reporting how the inferred threshold changes under alternative model choices (e.g., a power-law falloff).
  2. [Section 2.2.4 (local alignment and clipping)] The local alignment procedure (tenth-percentile zeroing, max-elevation centering, clipping, and rescaling) is applied independently to each evaluation region. This normalization changes the absolute elevation contrast before the CTF is computed, and it could systematically compress or enhance contrast as a function of building separation. The paper does not analyze the effect of this normalization on the fitted parameters A and sigma, nor on the resulting crossing distance. For instance, the centering between min and max of the reference data and the clipping to the test maximum elevation could create an apparent contrast falloff that mimics the Gaussian model even if the raw data do not exhibit that behavior. I ask the authors to quantify how much of the fitted sigma is attributable to the normalization versus the raw data, perhaps by running the pipeline with and without each alignment step on the synthetic data.
  3. [Section 2.2.4 and results (Figures 10, 12)] No uncertainty quantification is provided for any of the reported quantities. The fits in Figures 10 and 12 have no confidence intervals on A or sigma, and the reported threshold distances (2.5 m, 2 m) have no error bars. Given the large scatter visible in the CTF plots and the acknowledged variance sources in Section 2.3, it is not clear whether the difference between the Nellis and Jacksonville results is statistically significant or an artifact of the fit. Please report the number of evaluation regions used, bootstrapped confidence intervals on the fitted parameters and threshold distance, and standard goodness-of-fit diagnostics (e.g., residuals or R^2).
  4. [Section 2.2.4 and Figures 10, 12] The paper does not report the range of distances over which CTF data points are available, so it is unclear whether the C=0.2 crossing is an interpolation or an extrapolation of the fitted curve. If no building pairs with small separations exist at a site, the threshold distance is determined entirely by the assumed functional form beyond the data range. The authors should provide a histogram or density plot of evaluation-region distances and explicitly state whether the reported crossing lies within the observed distance range for each site and footprint source.
minor comments (5)
  1. [Figure 1 caption] There is a typo: "The test data data was generated" should be "The test data was generated."
  2. [Section 2.1] The sentence "The expected value of the CTF is not expected to be 1 as d approaches infinity" is redundant, and the model C(d) = A exp(...) implies that A is the large-distance limit, but the paper does not state whether A is constrained (e.g., positive or bounded).
  3. [Section 2.2.4] The local alignment description is difficult to follow without pseudo-code. Please provide a step-by-step algorithm or a numbered list to complement Figure 6.
  4. [Section 2.2.2] The construction of the binary building mask from the segmentation model relies on a confidence threshold, but the threshold value is not specified. Please state the value or indicate that it is user-adjustable.
  5. [Section 2.2 and results] The paper does not mention the computational cost or runtime of the pipeline, which would be useful for practitioners considering whether to use this method for large-area evaluations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: resolution values are transparently derived from a fitted model with independent synthetic validation, and self-citations are not load-bearing.

full rationale

The paper's central derivation is an adaptation of the contrast transfer function (CTF) to 3D elevation data. The model C(d)=A*exp(-(pi*sigma/d)^2) in Eq. (2) is explicitly introduced as a fitting function for the measured CTF-vs-distance scatter, and the reported threshold distance is read off from the fitted curve. This is a standard model-based summary of the data, not a prediction from a subset or a quantity defined in terms of itself. The paper states: 'This model will be used for determining a distance value for a given CTF threshold,' so the model's role is disclosed. The synthetic tribar experiment provides an external check: DSMs with known ground sample distances are progressively downsampled, and the fitted model's crossing points are compared with the known downsampling distances; this validation is independent of the real-site resolution claims. The local alignment and normalization steps in Section 2.2.4 are defined before the CTF calculation and are not constructed so that the output equals the input by design. Self-references [9] and [22] are cited for background and for other metrics (building-polygon RMSE and vertical-accuracy statistics); they are not load-bearing for the CTF methodology or its interpretation. No uniqueness theorem or ansatz is imported from the authors' prior work. Therefore, no circular step can be exhibited with the required specificity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central method rests on two fitted parameters (A and sigma) and several analyst-chosen thresholds, plus modeling assumptions about Gaussian behavior and building footprint fidelity. No new physical entities are introduced. The synthetic benchmark provides partial support for the model, but the real-world results are not independently verified against known resolution values.

free parameters (3)
  • sigma = not reported
    Fitted in Equation (2) to the CTF scatter; controls the width of the Gaussian model and determines the reported resolution distance. No value or confidence interval is given.
  • A = not reported
    Fitted in Equation (2) to model the asymptotic CTF limit as d approaches infinity, accounting for test/reference mismatch. Fit to the same data as sigma.
  • Evaluation region filtering thresholds = user-defined, unspecified
    Maximum centroid distance, maximum orthogonal distance, and reference CTF filter threshold are chosen by the user and affect which building pairs are included in the summary plot and fit. Exact values are not reported in the text.
assumptions (4)
  • ad hoc to paper The 3D data derivation process has a point-spread function that is approximately Gaussian, so CTF can be modeled as C(d)=A exp(-(pi sigma/d)^2).
    Introduced in Section 2.1, Equation (2), without derivation; motivated by the Fourier transform of a Gaussian PSF. Validated only on synthetic downsampled DSMs, not on real reconstruction processes.
  • domain assumption Building roof-to-ground elevation contrast, after local alignment, behaves analogously to intensity contrast in a tribar target.
    The CTF is computed from mean clipped elevations over building and center regions in Section 2.2.4. This requires buildings to be reasonably flat and adjacent ground to represent the zero level.
  • domain assumption The reference airborne lidar and derived DSM/DTM are sufficiently accurate and complete to serve as ground truth, and building footprints derived from segmentation or OSM represent true buildings.
    Sections 2.2.1 to 2.2.3; the authors note that incorrect class labels for larger buildings cause inaccurate polygon extraction, and transient structures in the reference data are not present in the test data.
  • domain assumption The system response is approximately the same in the x and y directions, so a single scalar CTF curve characterizes resolution.
    Section 2.1 states that the CTF is approximately a measure of the edge spread function under the assumption that the system response is the same in x and y.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Methods for evaluating the resolution of 3D data derived from satellite images." pith.science (2026). https://pith.science/paper/XX2QGD7F

@misc{pith2026250611876,
  author       = {Pith},
  title        = {Pith review of: Methods for evaluating the resolution of 3D data derived from satellite images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XX2QGD7F}},
  note         = {Machine review of arXiv:2506.11876}
}
read the original abstract

3D data derived from satellite images is essential for scene modeling applications requiring large-scale coverage or involving locations not accessible by airborne lidar or cameras. Measuring the resolution of this data is important for determining mission utility and tracking improvements. In this work, we consider methods to evaluate the resolution of point clouds, digital surface models, and 3D mesh models. We describe 3D metric evaluation tools and workflows that enable automated evaluation based on high-resolution reference airborne lidar, and we present results of analyses with data of varying quality.

Figures

Figures reproduced from arXiv: 2506.11876 by the authors.

Figure 1
Figure 1. A tribar reference dataset was created for analysis of the CTF methodology. Figures (a) - (d) are CTF [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. This example from the Nellis test site illustrates the 3DEP lidar point cloud, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Original OSM footprints13 (blue) and aligned OSM footprints (orange) over the reference digital surface model, the building mask derived from the lidar segmentation model, and the derived ground mask. Note that the ground mask has not had any morphological operations applied, while the building mask has. The reference DSM, building mask, and ground mask were derived from the 3DEP lidar point cloud14 over a small reg… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Aligned OSM footprints13 (orange) and segmentation model derived footprints (red) over the derived reference DSM and building mask derived from the lidar segmentation model. The reference DSM and building mask were derived from the 3DEP lidar point cloud14 over a small…
Figure 5
Figure 5. Figure 5: Example subregion of reference (a) and test (b) DSM over Nellis. Evaluation regions over the derived [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Figures (a)-(c) demonstrate the steps of the local alignment process. In all figures black represents the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Simple illustration of the calculation of CTF for an evaluation region. The black represents the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison of CTF values over evaluation regions for reference and test data. Red denotes the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: This illustration is an example of a CTF value below [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Example CTF summary plots over Nellis. Results in (a) and (b) were obtained using evaluation [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Google Earth visualization, Google© 2023 (a) of approximately 50 square kilometer region over Jacksonville, Florida for which example CTF example analysis has been completed for test DSM shown in (b). The test DSM was derived from MAXAR© satellite imagery, USGP-202110…
Figure 12
Figure 12. Figure 12: CTF summary plots are provided with respect to the evaluation regions displayed in Figure [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: The evaluation region centers over a reference DSM derived from 3DEP lidar data [PITH_FULL_IMAGE:figures/full_fig_p010_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    RPC STEREO PROCESSOR (RSP) – A SOFTW ARE PACKAGE FOR DIGITAL SURF ACE MODEL AND ORTHOPHOTO GENERATION FROM SATELLITE STEREO IMAGERY,

    Qin, R., “RPC STEREO PROCESSOR (RSP) – A SOFTW ARE PACKAGE FOR DIGITAL SURF ACE MODEL AND ORTHOPHOTO GENERATION FROM SATELLITE STEREO IMAGERY,” ISPRS An- nals of Photogrammetry, Remote Sensing and Spatial Information SciencesIII-1 (2016)

    work page 2016

  2. [2]

    Automatic 3d reconstruction from multi-date satellite images,

    Facciolo, G., De Franchis, C., and Meinhardt-Llopis, E., “Automatic 3d reconstruction from multi-date satellite images,” CVPR W(2017)

    work page 2017

  3. [3]

    Leveraging Vision Reconstruction Pipelines for Satellite Imagery,

    Zhang, K., Sun, J., and Snavely, N., “Leveraging Vision Reconstruction Pipelines for Satellite Imagery,” ICCV Workshop on 3D Reconstruction in the Wild(2019)

    work page 2019

  4. [4]

    Urban semantic 3d reconstruction from multiview satellite imagery,

    Leotta, M. J., Long, C., Jacquet, B., Zins, M., Lipsa, D., Shan, J., Xu, B., Li, Z., Zhang, X., Chang, S.-F., Purri, M., Xue, J., and Dana, K., “Urban semantic 3d reconstruction from multiview satellite imagery,” CVPR W(2019)

    work page 2019

  5. [5]

    Sat-NeRF: Learning multi-view satellite photogrammetry with transient objects and shadow modeling using RPC cameras,

    Mar ´ ı, R., Facciolo, G., and Ehret, T., “Sat-NeRF: Learning multi-view satellite photogrammetry with transient objects and shadow modeling using RPC cameras,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops(2022)

    work page 2022

  6. [6]

    Multi-date earth observation nerf: The detail is in the shadows,

    Mar ´ ı, R., Facciolo, G., and Ehret, T., “Multi-date earth observation nerf: The detail is in the shadows,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops(2023)

    work page 2023

  7. [7]

    BRDF-NeRF: Neural Radiance Fields with Optical Satellite Images and BRDF Modelling,

    Zhang, L., Rupnik, E., Nguyen, T. D., Jacquemoud, S., and Klinger, Y., “BRDF-NeRF: Neural Radiance Fields with Optical Satellite Images and BRDF Modelling,” arXiv (2024)

    work page 2024

  8. [8]

    FusionRF: High-Fidelity Satellite Neural Radiance Fields from Multispectral and Panchromatic Acquisitions,

    Sprintson, M., Chellappa, R., and Peng, C., “FusionRF: High-Fidelity Satellite Neural Radiance Fields from Multispectral and Panchromatic Acquisitions,” arXiv (2024)

    work page 2024

Show all 22 references
  1. [9]

    METRIC EV ALUATION PIPELINE FOR 3D MODELING OF URBAN SCENES,

    Bosch, M., Leichtman, A., Chilcott, D., Goldberg, H., and Brown, M., “METRIC EV ALUATION PIPELINE FOR 3D MODELING OF URBAN SCENES,” The International Archives of the Photogram- metry, Remote Sensing and Spatial Information SciencesXLII-1/W1 (2017)

    2017

  2. [10]

    Building3D: A Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds,

    Wang, R., Huang, S., and Yang, H., “Building3D: A Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds,” Proceedings of the IEEE/CVF International Conference on Computer Vision (2023)

    2023

  3. [11]

    Quantitative data quality metrics for 3d laser radar systems,

    Stevens, J. R., Lopez, N. A., and Burton, R. R., “Quantitative data quality metrics for 3d laser radar systems,” Laser Radar Technology and Applications XVI(2011)

    2011

  4. [12]

    Schott, J., Remote Sensing: The Image Chain Approach, Oxford University Press (2007)

    2007

  5. [13]

    Planet dump retrieved from https://planet.osm.org

    OpenStreetMap contributors, “Planet dump retrieved from https://planet.osm.org.” https://www. openstreetmap.org (2017)

    2017

  6. [14]

    3d elevation program (3dep) lidar data

    U.S. Geological Survey, “3d elevation program (3dep) lidar data.” The National Map (2021). Accessed March 15, 2024

    2021

  7. [15]

    SPU-Net: Self-Supervised Point Cloud Upsampling by Coarse- to-Fine Reconstruction With Self-Projection Optimization,

    Liu, X., Liu, X., Liu, Y.-S., and Han, Z., “SPU-Net: Self-Supervised Point Cloud Upsampling by Coarse- to-Fine Reconstruction With Self-Projection Optimization,” Trans. Img. Proc.31 (2022)

    2022

  8. [16]

    Pointcept: A Codebase for Point Cloud Perception Research

    Pointcept Contributors, “Pointcept: A Codebase for Point Cloud Perception Research.” https://github. com/Pointcept/Pointcept (2023)

    2023

  9. [17]

    DALES: A Large-scale Aerial LiDAR Data Set for Semantic Segmentation,

    Varney, N., Asari, V. K., and Graehling, Q., “DALES: A Large-scale Aerial LiDAR Data Set for Semantic Segmentation,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (2020)

    2020

  10. [18]

    STPLS3D: A Large-Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset,

    Chen, M., Hu, Q., Yu, Z., THOMAS, H., Feng, A., Hou, Y., McCullough, K., Ren, F., and Soibelman, L., “STPLS3D: A Large-Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset,” 33rd British Machine Vision Conference(2022)

    2022

  11. [19]

    swissSURF ACE3D

    Federal Office of Topography swisstopo, “swissSURF ACE3D.” https://www.swisstopo.admin.ch/en/ height-model-swisssurface3d. Accessed: 2025-01-21

    2025

  12. [20]

    Algorithms for the reduction of the number of points required to repre- sent a digitized line or its caricature,

    Douglas, D. H. and Peucker, T. K., “Algorithms for the reduction of the number of points required to repre- sent a digitized line or its caricature,” Cartographica: The International Journal for Geographic Information and Geovisualization 10 (1973)

    1973

  13. [21]

    Polygonal building extraction by frame field learning,

    Girard, N., Smirnov, D., Solomon, J., and Tarabalka, Y., “Polygonal building extraction by frame field learning,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2021)

    2021

  14. [22]

    Cumulative assessment for urban 3d modeling,

    Hagstrom, S., Pak, H. W., Ku, S., Wang, S., Hager, G., and Brown, M., “Cumulative assessment for urban 3d modeling,” 2021 IEEE International Geoscience and Remote Sensing Symposium(2021)

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.