acceptodds
Withdrawn from ICLR 2027

GaussLift: Renderer-Consistent DSM Export from Satellite Gaussian Fields

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D Gaussian Splatting (3DGS) can reproduce satellite imagery with high photometric fidelity, yet a learned Gaussian field does not uniquely determine a metrically accurate digital surface model (DSM). The ambiguity arises at the representation–measurement interface: primitive centers, projected footprints, and renderer-visible support can induce different elevation surfaces from the same field. We introduce GaussLift, which makes this interface explicit by reusing front-to-back alpha-compositing weights to define a normalized renderer-conditioned altitude observable, then localizing supported samples in world coordinates and realizing them on the DSM grid. The same observable is used for surface supervision and export, without introducing an auxiliary DSM decoder. Fixed-field controls on DFC2019 confirm that changing only the export operator alters both reconstruction error and coverage; under a common-mask comparison across frozen fields, renderer-conditioned altitude attains the lowest MAE and RMSE among the tested readouts. Across four scene-specific reconstructions, the final 12-member map-space ensemble achieves 0.787 m raw MAE, 2.344 m RMSE, and 95.91% coverage. The underlying Gaussian fields remain renderable from 68 held-out views without target-view fitting, reaching 22.562 dB mean PSNR. These results show that surface readout is a consequential component of metric 3D Gaussian reconstruction, complementary to representation learning.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Satellite Imagery, Digital Surface Model, Remote Sensing, Neural Rendering, 3D Reconstruction, Gaussian-to-DSM, Geospatial Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6AmQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GaussLift converts a satellite Gaussian field into a georeferenced DSM. It computes altitude using the RGB renderer's visibility weights, maps the resulting samples into world coordinates, and aggregates them on a grid. Each experimental field is fitted to one RGB image and the complete reference DSM. The paper compares export rules on frozen fields and reports both individual exports and a twelve-member DSM ensemble.

Critical strengths
  1. The paper evaluates export accuracy together with coverage. Table 3(a) shows why errors over different valid regions cannot directly establish which exporter is better. Table S2 adds a comparison on common valid cells. This is a useful diagnostic for evaluating Gaussian-derived DSMs.

  2. The fitted fields retain a rendering capability. The paper evaluates individual fields on 68 held-out RGB views without fitting to those views. It also clearly distinguishes these rendering results from the fused DSM and discloses the use of full reference-DSM supervision.

Critical weaknesses
  1. The main experiment fits a known DSM and then measures how well it recovers that same DSM. Section 4.1 uses one satellite image and the complete reference DSM to fit each Gaussian field. The exported DSM is then evaluated against that same reference. Thus, the DSM error measures how well supplied geometry survives fitting and export; it does not measure recovery of unknown geometry from satellite imagery. The authors explicitly limit the task to scene fitting, but its practical value still needs evidence. If a complete DSM is already available, what does this pipeline gain over rendering a textured DSM or mesh? The direct height-prediction baselines do not answer that question.

  2. The exporter is tested only on fields trained to favor its own output. Equations (17)-(18) directly supervise the altitude statistic that GaussLift later exports. Freezing the field makes the export comparison controlled, but the field has already been optimized for this statistic. Table S2 establishes a benefit on these fields: common-mask MAE improves from 0.794 to 0.785 m, and RMSE from 2.403 to 2.349 m. The MAE gain is only 9 mm across four scenes. Testing independently reconstructed fields, or crossing training losses with export rules, would show whether the exporter offers a useful advantage beyond this specific training setup.

  3. The strongest system result lacks a baseline comparison at the same computational budget. Table 3(b) reduces MAE from 0.907 to 0.787 m by fusing twelve independently optimized exports. The authors correctly attribute this improvement to fusion. However, Table 1 and Table S1 do not report an equivalent ensemble or establish a matched compute budget for the raster baselines. Without fusion, GaussLift has lower MAE than LUMNet but slightly higher RMSE: 2.585 versus 2.545 m. Consequently, Appendix D.1 overstates the evidence when it says the comparison isolates the intermediate representation. The twelve member configurations, total cost, and matched-budget results are needed to assess the system's advantage.

Questions for authors
  1. Can GaussLift improve DSM export from satellite Gaussian fields reconstructed without the complete evaluation DSM and without its proposed altitude loss? This is the most important missing experiment for assessing the exporter.

  2. If full-DSM scene fitting is the intended application, how does the method compare with a textured DSM or mesh in held-out rendering quality, storage, and computation?

  3. What are the twelve ensemble members' resolutions and training configurations? Does the advantage remain when the raster baselines receive a comparable fitting and fusion budget?

Reviewer_rhrpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the export of georeferenced digital surface models (DSMs) from satellite 3D Gaussian Splatting (3DGS) fields. It addresses the distinction between a renderable Gaussian field and a single-valued elevation surface on a geographic grid. GaussLift computes renderer-conditioned altitude from normalized alpha-compositing weights, rejects pixels with insufficient support, and localizes the supported samples through an RPC-derived affine camera. The samples are rasterized onto the DSM grid by retaining the maximum valid height in each cell. Each scene is optimized using one RGB image and the complete reference DSM, with the altitude loss directly supervising the statistic used for export. Experiments on four DFC2019 scenes compare against published raster-based DSM predictors, protocol-matched raster baselines, and alternative readouts of frozen Gaussian fields. The final 12-member ensemble achieves 0.787 m MAE, 2.344 m RMSE, and 95.91% coverage under this scene-specific, full-DSM-supervision protocol.

Critical strengths
  1. The paper clearly identifies a consequential distinction between rendering and surface extraction. A Gaussian field represents volumetric contributions and visibility, whereas a DSM requires one elevation per geographic grid cell. Good image rendering therefore does not by itself establish that the exported DSM is geometrically accurate. Making this distinction explicit gives the work a useful focus for satellite reconstruction and geospatial applications.

  2. Treating export as a first-class object of study is valuable. Freezing an optimized field and changing only the export operator allows the consequences of different export rules to be examined without retraining the representation. This is a useful experimental design, even though its conclusions remain conditional on how the field was trained. Separately listing published cross-scene results and protocol-matched retraining results in Table 1, together with the explicit scope statement in §4.1, also helps readers interpret the evidence appropriately.

Critical weaknesses
  1. The technical advance over existing satellite-Gaussian elevation rendering is not sufficiently established. EOGS (CVPR 2025, Eq. (7)) already renders metric elevation using Gaussian alpha-compositing weights. This is closely related to the altitude statistic in the submitted Eq. (7), although EOGS's displayed expression is a weighted sum rather than an explicitly support-normalized moment. The paper needs to explain which differences in normalization, support handling, supervision, localization, and grid realization constitute its contribution, and where their quantitative benefits appear. Table 1 also lacks a direct comparison with an existing satellite-3DGS DSM pipeline. Together, these omissions make it difficult to assess the significance of the proposed export formulation relative to the closest prior work.

  2. The experiments do not sufficiently separate the effects of the readout, its supervision, and ensemble aggregation. The altitude loss, L_alt in Eq. (18), directly supervises the renderer-conditioned altitude A_\Pi used by GaussLift, while the center- and footprint/opacity-based alternatives do not receive comparable supervision. Freezing this field shows that export choices matter for that trained representation; it does not establish that the proposed readout is intrinsically better. The common-mask MAE comparison in Table S2 is 0.785 versus 0.794 m, so the extent to which this advantage persists under other training conditions matters to the central claim.

Questions for authors
  1. How does the submitted Eq. (7) differ from EOGS's elevation-rendering Eq. (7), both mathematically and in implementation? Which benefits can be attributed specifically to normalization and support handling, and is there a direct satellite-3DGS DSM comparison under a compatible protocol that establishes the contribution beyond this precedent?
  2. Can you provide a small field-by-readout comparison that includes a field trained without L_alt, evaluated on matched support? Does the readout ordering in Table S2 persist, and how much of the 0.785 versus 0.794 m common-mask MAE difference is attributable to the supervision of A_Pi?
Reviewer_uYyYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper demonstrates how different export rules produce different digital surface models (DSMs) from the same frozen satellite Gaussian field. It proposes GaussLift, which computes a normalized altitude moment using front-to-back alpha-compositing weights, filters samples by accumulated opacity, localizes them through an RPC-derived affine camera, and selects the maximum height within each occupied map cell. Each scene is fitted using one RGB image and the complete reference DSM, with the same altitude observable used for supervision and export. The final DSM is obtained by an ensemble, constructed with a cell-wise median fusion of 12 independent DSM exports. Experiments on four DFC2019 scenes show reconstruction accuracy, export-rule dependency, and rendering from held-out viewpoints without additional tuning.

Critical strengths
  • The paper identifies a useful distinction between learning a Gaussian representation and extracting a DSM from it. Experiments with the proposed Gaussian fields (Tables 3 (a) and S2) demonstrate changing the exporter while keeping the fields fixed can alter both reconstruction error and coverage.
Critical weaknesses
  • My largest concern on this paper is that the same reference DSM is used to both train the 3D Gaussian field and evaluate the exported DSM (Line 318). More precisely, the Lalt\mathcal{L}_{\text{alt}} is supervised directly by the reference DSM, as described in Equation (18), and Gaussian initialization samples points near valid reference-DSM surfaces, as described in Supplementary Section B.

    • Other than novel view synthesis, the practical benefit of converting an already-known DSM into this field representation also needs a stronger demonstration.

    • The paper does not establish whether the proposed readout remains advantageous for fields learned under different objectives. Evaluating independently trained Gaussian fields would better establish the generality of the DSM export contribution.

  • The DSM exporter is evaluated only on four Jacksonville scenes and one field-construction pipeline. Additional calibrated satellite scenes, such as IARPA 2016 (as evaluated by the cited EOGS++), and fields produced by another reconstruction method would establish whether the observed accuracy-coverage behavior persists beyond this setting.

Questions for authors
  • Please address the issues raised in Weaknesses.

  • Could the authors separately characterize sensitivity to reduced or noisy DSM supervision? What is the performance change when λalt\lambda_{\text{alt}} is set to zero?

  • In Table 1, how are the reproduced CMT, HeightFormer, and LUMNet initialized and optimized? Please specify their trainable components, optimization budgets, and stopping criteria, and provide evidence of convergence. Could extended training budget change the conclusion of Table 1?

Suggestions:

  • The RGB images in rows 2 and 3 in Figure 3 appear to be incorrectly rotated, compared to the visibility, altitude, and DSM visualizations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.