Pith. sign in

REVIEW 4 major objections 6 minor 83 references

SkyLume isolates illumination from viewpoint in 100K+ urban aerial images, showing current 3D reconstruction methods are not robust to sunlight changes.

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 16:13 UTC pith:T2SPAVQ5

load-bearing objection SkyLume's dataset is the real contribution and worth serious refereeing; the static-scene assumption is unquantified, TCC is self-referential, and there's no release link yet. the 4 major comments →

arxiv 2512.14200 v3 pith:T2SPAVQ5 submitted 2025-12-16 cs.CV

Beyond a Single Light: A Large-Scale Aerial Dataset for Urban Scene Reconstruction Under Varying Illumination

Zhuoxiao Li , Wenzong Ma , Taoyu Wu , Jinjing Zhu , Shuai Zhang , Jing OU , Tongyan Hua , Yinrui Ren
show 3 more authors
Rongjun Qin Hui Xiong Wufan Zhao
This is my paper
classification cs.CV
keywords UAV datasetillumination variation3D reconstructionGaussian splattinginverse renderingalbedo consistencyLiDAR ground truthnovel view synthesis
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 introduces SkyLume, a large-scale real-world aerial dataset that photographs the same ten urban regions at three times of day—morning, noon, and late afternoon—with LiDAR ground truth, registered camera poses, and per-frame depth and normals. The central claim is that this is the first benchmark that isolates illumination change from viewpoint change at city scale, so differences between captures can be attributed to lighting rather than to camera motion. Using this resource, the authors show that current Gaussian-splatting reconstruction methods (a technique that models scenes as clouds of 3D ellipsoids) degrade noticeably under strong sunlight: shadows get baked into geometry, albedo estimates drift across time, and rendering quality drops on facades, glass, and water. They also introduce a metric, the Temporal Consistency Coefficient (TCC), that scores cross-time albedo stability, turning multi-temporal robustness into a measurable target.

Core claim

SkyLume consists of more than 100,000 six-megapixel images from five synchronized views over ten urban regions, each region flown in the morning, at noon, and in the late afternoon along identical RTK-guided waypoints. LiDAR scans provide metric ground truth, and a unified structure-from-motion registration locks all time slots into one coordinate system, so the only intended variable is illumination. Benchmarked 3D Gaussian Splatting variants show substantial quality loss under direct sun: cast shadows and moving penumbras are frequently reconstructed as solid geometry, albedo from inverse rendering retains shading and varies across time, and novel view synthesis blurs on weakly textured an

What carries the argument

The load-bearing mechanism is the capture protocol: repeating the same flight trajectory at three times of day and registering all images to a common LiDAR-guided coordinate system, which converts multi-temporal capture into a controlled illumination experiment. The TCC metric is the evaluation mechanism: it renders albedo from K fixed test viewpoints in each of the three time slots, compares each slot against the temporal mean, and combines MAE, RMSE, SSIM, and LPIPS into a single [0,1] consistency score.

Load-bearing premise

Each urban scene is effectively static across the three capture sessions, so measured differences are attributed to illumination rather than to scene change.

What would settle it

Compare TCC scores and geometry F1 in regions with known moving objects (vehicles, pedestrians, cloud-shadow boundaries) versus static regions; if dynamic regions show systematically higher instability, the attribution to illumination alone fails.

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

If this is right

  • Current 3D Gaussian Splatting methods can now be ranked and improved on a controlled illumination-robustness benchmark, rather than on ad hoc captures.
  • Researchers can isolate shadow-induced geometry bias by comparing reconstructions from sunlit versus overcast slots against the same LiDAR ground truth.
  • The provided solar geometry (elevation and azimuth per image) enables future de-shadowing, relighting, and illumination-aware training.
  • The standardized splits allow direct comparison of novel view synthesis methods under identical viewpoint and lighting conditions.
  • City-scale inverse rendering can be evaluated for material/light disentanglement via TCC, a task previously limited to small objects or synthetic scenes.

Where Pith is reading between the lines

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

  • The static-scene assumption is the clear risk: urban areas contain moving vehicles, pedestrians, and cloud shadows, so TCC and cross-time geometry scores could conflate illumination effects with object dynamics; masking dynamic regions would sharpen the benchmark.
  • The TCC design—fixed viewpoints, temporal mean, per-viewpoint aggregation—is general enough to be applied to other repeat-photography setups, such as ground-level cameras or multi-day satellite revisits, as a measure of appearance stability.
  • The observed result that diffuse illumination yields the most stable geometry suggests a practical capture policy: fly under overcast conditions when geometry is the priority, and use sunlit slots for appearance and albedo studies.
  • The dataset could double as a real-world domain-shift test for generalizable reconstruction models trained on synthetic data, since it provides true illumination variation with ground truth.

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. SkyLume is introduced as a large-scale, real-world UAV aerial dataset for urban scene reconstruction under varying illumination. The dataset comprises 10 urban regions, more than 100K six-megapixel images captured from five directions (four oblique and one nadir) at three times of day (morning, noon, afternoon), with per-scene LiDAR scans, unified 6-DoF poses, LiDAR-derived ground-truth depth and normals, and solar-geometry annotations. The paper also proposes Temporal Consistency Coefficient (TCC), a metric intended to evaluate cross-time albedo stability for inverse rendering, and benchmarks several 3D Gaussian Splatting variants on geometry, novel view synthesis, and albedo-consistency tracks. The central claim is that SkyLume is the first real-world aerial dataset that systematically isolates illumination changes from viewpoint changes at city scale, enabling rigorous evaluation of illumination-robust reconstruction.

Significance. If the central claims hold, SkyLume fills a genuine gap in aerial 3D vision: prior real-world UAV datasets either do not revisit the same area under systematically varying illumination or lack high-precision geometry. The dataset's acquisition is carefully described, and the reported alignment statistics (median reprojection error 0.70 px, millimeter-level relative pose uncertainty, Table 7) suggest a high-quality, unified multi-temporal SfM solution. The per-frame LiDAR depth and solar-geometry annotations add practical value for downstream tasks. The benchmark experiments reveal plausible trends, such as geometry degradation under hard shadows. However, the evaluation methodology has two load-bearing weaknesses: (1) the static-scene assumption under which illumination is isolated is asserted but not quantified or enforced; (2) the TCC metric measures consistency against the temporal mean of the very albedo maps being evaluated, so it cannot validate whether the disentangled albedo is correct, and it rewards trivially constant albedo. These issues do not necessarily invalidate the dataset itself, but they currently weaken the claimed contributions. With substantial revisions—parti

major comments (4)
  1. [§3.2, §6.1, Table 6, Fig. 6 caption] yes
  2. [§4.1, Eq. (8)-(14)] yes
  3. [§3.3, Fig. 3, §3.2] yes
  4. [General] yes
minor comments (6)
  1. [§4.3, Table 3] The text says 'Period 0 and Period 1' (Sec. 4.3) but Table 3 lists 'Period 1' and 'Period 2'. This inconsistency should be corrected. Also, 'between the the meshes' is a typo.
  2. [Eq. (10)-(11)] The constants in the saturation functions are not justified. Units of MAE and RMSE should be stated (e.g., 0-1 albedo range), and the choice of 10 should be discussed.
  3. [Table 1] Typo 'ISPRS-Bencnmark' should be 'ISPRS-Benchmark'. Also, the 'Light' column header could be clarified as 'Varying Illumination'.
  4. [Fig. 6 caption] The phrase 'differences arise solely from illumination' is too strong given the static-scene issue; it should be rephrased to match the more cautious claim in the supplementary.
  5. [References] References [11] and [12] are duplicates of the same 3DGS paper. Please consolidate.
  6. [§3.3] Typo 'RealityScanto' appears in Sec. 3.3; also 'COLMAP' formatting is inconsistent.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The dataset itself is an empirical resource with no derivation. The main epistemic costs are the static-scene assumption, the use of proprietary LiDAR-guided meshing as ground truth, and the self-referential TCC metric. The TCC weights and error-saturation factor are free parameters affecting all inverse-rendering conclusions.

free parameters (4)
  • TCC component weights = α=β=θ=0.2, γ=0.4
    Manually chosen in Sec. 4.1; no optimization or justification is given for this particular weighting.
  • TCC MAE/RMSE saturation scale = 10
    Equation (10)-(11) map errors to [0,1] via 1 - clip(10*error); this factor is chosen by hand and controls the score distribution.
  • Geometry distance thresholds = 0.25, 0.5, 0.75 m
    Fixed thresholds for F1 precision/recall; reasonable but arbitrary and not derived from LiDAR noise estimates.
  • Unified single-GPU training schedule = 90k iterations, densification interval 300, opacity reset 9000, densification disabled after 60k
    Chosen by the authors to make comparisons fair under a single 80GB GPU; not a property of the dataset, but affects all benchmark conclusions.
axioms (4)
  • domain assumption Urban scenes are static across the three time-of-day captures
    Sec. 3.2 and Supp. Sec. 6.1 state that differences across slots predominantly reflect illumination rather than changes in scene structure. Moving vehicles, pedestrians, vegetation and changing cloud shadows are not filtered or quantified.
  • domain assumption LiDAR-guided SfM registration in RealityScan yields cm-level accurate poses
    Sec. 3.3 reports mm-level relative camera position uncertainty, but the pipeline relies on proprietary software and manual GCP corrections, and the absolute accuracy against independent checkpoints is not fully reported in the main text.
  • domain assumption LiDAR-guided MVS mesh is a valid ground-truth surface
    Sec. 3.3 constructs the GT mesh by fusing the same RGB imagery with LiDAR in RealityScan and manually repairing water surfaces. The mesh is then used as ground truth to evaluate methods trained on the same RGB imagery.
  • domain assumption The 80/60 overlap, 120 m altitude, and 1 Hz trigger provide sufficient multiview coverage
    Sec. 3.2 states these parameters follow standard photogrammetric practice; no ablation is given to show they are sufficient for the claimed GT quality across all 10 regions.
invented entities (1)
  • Temporal Consistency Coefficient (TCC) no independent evidence
    purpose: Evaluate cross-time albedo stability and 'directly evaluate the robustness of the disentanglement of light and material'.
    TCC is computed against the temporal mean of the method's own albedo outputs; there is no external albedo ground truth. A degenerate method that always outputs a constant albedo would receive a perfect TCC score, so the metric does not independently validate material/light disentanglement.

pith-pipeline@v1.3.0-alltime-deepseek · 17687 in / 10525 out tokens · 85330 ms · 2026-08-03T16:13:00.604043+00:00 · methodology

0 comments
read the original abstract

Recent advances in Neural Radiance Fields and 3D Gaussian Splatting have demonstrated strong potential for large-scale UAV-based 3D reconstruction tasks by fitting the appearance of images. However, real-world large-scale captures are often based on multi-temporal data capture, where illumination inconsistencies across different times of day can significantly lead to color artifacts, geometric inaccuracies, and inconsistent appearance. Due to the lack of UAV datasets that systematically capture the same areas under varying illumination conditions, this challenge remains largely underexplored. To fill this gap, we introduceSkyLume, a large-scale, real-world UAV dataset specifically designed for studying illumination robust 3D reconstruction in urban scene modeling: (1) We collect data from 10 urban regions data comprising more than 100k high resolution UAV images (four oblique views and nadir), where each region is captured at three periods of the day to systematically isolate illumination changes. (2) To support precise evaluation of geometry and appearance, we provide per-scene LiDAR scans and accurate 3D ground-truth for assessing depth, surface normals, and reconstruction quality under varying illumination. (3) For the inverse rendering task, we introduce the Temporal Consistency Coefficient (TCC), a metric that measuress cross-time albedo stability and directly evaluates the robustness of the disentanglement of light and material. We aim for this resource to serve as a foundation that advances research and real-world evaluation in large-scale inverse rendering, geometry reconstruction, and novel view synthesis.

Figures

Figures reproduced from arXiv: 2512.14200 by Hui Xiong, Jing Ou, Jinjing Zhu, Rongjun Qin, Shuai Zhang, Taoyu Wu, Tongyan Hua, Wenzong Ma, Wufan Zhao, Yinrui Ren, Zhuoxiao Li.

Figure 1
Figure 1. Figure 1: We present SkyLume, the first comprehensive real-world UAV dataset centered on illumination variation. It provides 6K￾resolution five-direction imagery from three daily captures along identical RTK-guided flight paths, paired with LiDAR-derived ground truth including precise meshes and per-frame depth maps and normal maps under unified 6-DoF poses. The benchmark enables rigorous evaluation of 3D reconstruc… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset collection and processing pipeline. (a) A survey-grade UAV stack flies the same RTK-guided route at three times of day to capture five-direction 6K imagery and LiDAR. (b) A unified LiDAR-guided SfM registration includes three periods and refines poses by point-rendering LiDAR into the cameras. (c) A LiDAR-guided MVS to produce an high-quality aligned ground-truth geometry. (d) We release per-period… view at source ↗
Figure 3
Figure 3. Figure 3: Geometry post-processing. We first build geometry ground truth via LiDAR-guided MVS. For reflective area such as river and lake, we manually repair water surfaces to correct MVS failures and ensure geometric continuity. 3.2. Data Collection As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the ground-truth models. We visualize post-processed meshes for six representative medium-scale scenes. fusing the three-time-slot RGB imagery with the LiDAR data. The LiDAR provides a metric scaffold that regularizes depth in weak-texture and shadowed regions. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Albedo and TCC visualization across two time slots. Note that Period 2 is rendered from the viewpoints of Period 1 to fix camera pose, so differences arise solely from illumination. For detailed implementation information, please refer to the supplementary material. 4.1. Benchmark Metrics Illumination robustness. In real-world conditions, it is not feasible to capture accurate ground-truth albedo. Therefor… view at source ↗
Figure 7
Figure 7. Figure 7: Geometry visualization. Top: Geometry for Gym and Residence, red and yellow frames are the zoomed-in surface normals. Bottom: Single Period 1 mesh comparison. Under the sunlit Period 1, all methods exhibit holes and breakups. pairwise consistency F-1 score between the the meshes in Tab. 3. Two lighting regimes are highlighted in [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional Visualization of the Impact of Strong Il￾lumination. Cast shadows break multi-view photometric consis￾tency, yielding shadow-as-geometry artifacts. In reflective planes, view-dependent appearance leads to blurred textures and detail loss in NVS. scenes: geometry-oriented pipelines incur a noticeable loss in photometric fidelity and therefore lag in rendering qual￾7 [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of novel view synthesis under Period 1 (sunlit). We visualize renderings from shadow-dominated regions to highlight some baselines lost details. ity. Within the NVS group, Abs-GS [56] is the most reliable under UAV oblique viewpoints, and Octree-GS [32] recov￾ers fine structures. By contrast, 2DGS [8] produces blurred renderings in all four representative scenes, as illustrated in [PITH_FULL… view at source ↗
Figure 10
Figure 10. Figure 10: The DJI M350 RTK UAV platform equipped with a [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Challenging cases across our dataset. We show representative crops for four typical difficulty patterns: (a) water bodies with strong specular reflections and weak texture, (b) glass fac¸ades with transparency and mirror-like reflections, (c) complex architectures with intricate geometry and self-occlusions, and (d) high-density urban blocks with tightly packed buildings [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 12
Figure 12. Figure 12: Reconstruction challenges under partly cloudy illumination. In partly cloudy scenes, moving clouds create strongly varying local illumination, so the same region appears sunlit in some views and shadowed in others. This breaks photometric consistency across viewpoints and leads to color inconsistencies. (Glass Facade). Across all scenes, the total number of images per scene ranges from roughly 5k to over … view at source ↗
Figure 14
Figure 14. Figure 14: Effect of the modified pruning schedule. Qualita￾tive comparison after the first opacity reset iteration (9000) and pruning. left: Opacity reset iteration is 9,000, densify interval is 300, and we visualized the result of 9,301 (9000+300+1) iter￾ation; Right: Opacity reset iteration is 9,000, image count and densify interval is 1770, and we visualized the result of 10771 (9000+1770+1) iteration. Algorithm… view at source ↗
Figure 15
Figure 15. Figure 15: Additional geometry ground-truth results. We visualize the geometry ground truth of the Town scene, which features dense high-rise blocks, narrow streets, and deeply occluded courtyards. Our LiDAR-guided MVS pipeline produces a metrically accurate and topologically complete reference surface that preserves fine fac¸ade details and inner-structure layout, providing a reliable ground-truth benchmark for eva… view at source ↗
Figure 16
Figure 16. Figure 16: Additional bird’s-eye visualizations rendered with the SuperSplat web viewer. [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Additional TCC-albedo visualization of Ref-GS. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Additional TCC-albedo visualization of Ref-Gaussian. [PITH_FULL_IMAGE:figures/full_fig_p019_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Additional TCC-albedo visualization of GS-IR. [PITH_FULL_IMAGE:figures/full_fig_p020_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Qualitative comparison between SOTA monocular depth/normal estimators on SkyLume dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Additional bird’s-eye geometry visualizations. [PITH_FULL_IMAGE:figures/full_fig_p021_21.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

83 extracted references · 11 linked inside Pith

  1. [1]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. 2024. 3, 5, 7

    2024

  2. [2]

    Gigags: Scaling up planar-based 3d gaussians for large scene surface reconstruction, 2024

    Junyi Chen, Weicai Ye, Yifan Wang, Danpeng Chen, Di Huang, Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. Gigags: Scaling up planar-based 3d gaussians for large scene surface reconstruction, 2024. 2, 3

    2024

  3. [3]

    Dogaussian: Distributed- oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus, 2024

    Yu Chen and Gim Hee Lee. Dogaussian: Distributed- oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus, 2024. 2

    2024

  4. [4]

    Firestereo: Forest infrared stereo dataset for uas depth perception in visually degraded environments.arXiv preprint arXiv:2409.07715, 2024

    Devansh Dhrafani, Yifei Liu, Andrew Jong, Ukcheol Shin, Yao He, Tyler Harp, Yaoyu Hu, Jean Oh, and Sebastian Scherer. Firestereo: Forest infrared stereo dataset for uas depth perception in visually degraded environments.arXiv preprint arXiv:2409.07715, 2024. 3

    Pith/arXiv arXiv 2024

  5. [5]

    Mini-splatting: Represent- ing scenes with a constrained number of gaussians, 2024

    Guangchi Fang and Bing Wang. Mini-splatting: Represent- ing scenes with a constrained number of gaussians, 2024. 3

    2024

  6. [6]

    Mid-air: A multi-modal dataset for extremely low altitude drone flights

    Michael Fonder and Marc Van Droogenbroeck. Mid-air: A multi-modal dataset for extremely low altitude drone flights. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition workshops, pages 0–0, 2019. 3

    2019

  7. [7]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh recon- struction and high-quality mesh rendering.arXiv preprint arXiv:2311.12775, 2023

    Antoine Gu ´edon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh recon- struction and high-quality mesh rendering.arXiv preprint arXiv:2311.12775, 2023. 3

    Pith/arXiv arXiv 2023

  8. [8]

    2d gaussian splatting for geometrically accu- rate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accu- rate radiance fields. InSIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 3, 5, 7, 8

    2024

  9. [9]

    Horizon- gs: Unified 3d gaussian splatting for large-scale aerial-to- ground scenes

    Lihan Jiang, Kerui Ren, Mulin Yu, Linning Xu, Junting Dong, Tao Lu, Feng Zhao, Dahua Lin, and Bo Dai. Horizon- gs: Unified 3d gaussian splatting for large-scale aerial-to- ground scenes. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26789–26799, 2025. 2, 3

    2025

  10. [10]

    Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces

    Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xi- aoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 5322–5332, 2024. 3

    2024

  11. [11]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023. 5

    2023

  12. [12]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  13. [13]

    3d gaussian splat- ting as markov chain monte carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splat- ting as markov chain monte carlo. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 3

    2024

  14. [14]

    Shading-aware multi-view stereo

    Fabian Langguth, Kalyan Sunkavalli, Sunil Hadap, and Michael Goesele. Shading-aware multi-view stereo. In European Conference on Computer Vision, pages 469–485. Springer, 2016. 2

    2016

  15. [15]

    SGLoc: Scene geometry en- coding for outdoor LiDAR localization

    Wen Li, Shangshu Yu, Cheng Wang, Guosheng Hu, Siqi Shen, and Chenglu Wen. SGLoc: Scene geometry en- coding for outdoor LiDAR localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9286–9295, 2023. 3

    2023

  16. [16]

    Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond

    Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhen- zhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3205–3215, 2023. 2, 3

    2023

  17. [17]

    Garcia- Fernandez, Yong Yue, Eng Gee Lim, and Xiaohui Zhu

    Zhuoxiao Li, Shanliang Yao, Yijie Chu, Angel F. Garcia- Fernandez, Yong Yue, Eng Gee Lim, and Xiaohui Zhu. Mvg-splatting: Multi-view guided gaussian splatting with adaptive quantile-based geometric consistency densification,

  18. [18]

    Garc´ıa-Fern´andez, Andrew Levers, Jason Ralph, and Xiaohui Zhu

    Zhuoxiao Li, Shanliang Yao, Taoyu Wu, Yong Yue, Wu- fan Zhao, Rongjun Qin, ´Angel F. Garc´ıa-Fern´andez, Andrew Levers, Jason Ralph, and Xiaohui Zhu. Ulsr-gs: Urban large- scale surface reconstruction gaussian splatting with multi- view geometric consistency.ISPRS Journal of Photogram- metry and Remote Sensing, 230:861–880, 2025. 3

    2025

  19. [19]

    Gs-ir: 3d gaussian splatting for inverse rendering

    Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 3, 5, 6

    2024

  20. [20]

    Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y . Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647, 2025. 5

    Pith/arXiv arXiv 2025

  21. [21]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiy- ong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5166–5175, 2024. 2, 3

    2024

  22. [22]

    Capturing, reconstructing, and simulating: the urbanscene3d dataset

    Liqiang Lin, Yilin Liu, Yue Hu, Xingguang Yan, Ke Xie, and Hui Huang. Capturing, reconstructing, and simulating: the urbanscene3d dataset. InEuropean Conference on Computer Vision, pages 93–109. Springer, 2022. 2, 3

    2022

  23. [23]

    Urbanscene3d: A large scale urban scene dataset and simulator

    Yilin Liu, Fuyou Xue, and Hui Huang. Urbanscene3d: A large scale urban scene dataset and simulator. 2021. 2, 3

    2021

  24. [24]

    Citygaus- sian: Real-time high-quality large-scale scene rendering with gaussians.arXiv preprint arXiv:2404.01133, 2024

    Yang Liu, He Guan, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. Citygaus- sian: Real-time high-quality large-scale scene rendering with gaussians.arXiv preprint arXiv:2404.01133, 2024. 3

    Pith/arXiv arXiv 2024

  25. [25]

    Citygaussianv2: Efficient and geometri- cally accurate reconstruction for large-scale scenes

    Yang Liu, Chuanchen Luo, Zhongkai Mao, Junran Peng, and Zhaoxiang Zhang. Citygaussianv2: Efficient and geometri- cally accurate reconstruction for large-scale scenes. InICLR,

  26. [26]

    Scaffold-gs: Structured 3d 9 gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d 9 gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654–20664, 2024. 3

    2024

  27. [27]

    Switch-nerf: Learning scene de- composition with mixture of experts for large-scale neural radiance fields

    Zhenxing Mi and Dan Xu. Switch-nerf: Learning scene de- composition with mixture of experts for large-scale neural radiance fields. InInternational Conference on Learning Representations (ICLR), 2023. 2

    2023

  28. [28]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2

    2021

  29. [29]

    F. Nex, M. Gerke, F. Remondino, H.-J. Przybilla, M. B¨aumker, and A. Zurhorst. Isprs benchmark for multi- platform photogrammetry.ISPRS Annals of the Photogram- metry, Remote Sensing and Spatial Information Sciences, II- 3/W4:135–142, 2015. 2, 3

    2015

  30. [30]

    Ntu viral: A visual- inertial-ranging-lidar dataset, from an aerial vehicle view- point.The International Journal of Robotics Research, 41 (3):270–280, 2022

    Thien-Minh Nguyen, Shenghai Yuan, Muqing Cao, Yang Lyu, Thien H Nguyen, and Lihua Xie. Ntu viral: A visual- inertial-ranging-lidar dataset, from an aerial vehicle view- point.The International Journal of Robotics Research, 41 (3):270–280, 2022. 3

    2022

  31. [31]

    Multi-view relighting using a geometry-aware network.ACM Trans

    Julien Philip, Micha ¨el Gharbi, Tinghui Zhou, Alexei A Efros, and George Drettakis. Multi-view relighting using a geometry-aware network.ACM Trans. Graph., 38(4):78–1,

  32. [32]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024. 3, 5, 8

    Pith/arXiv arXiv 2024

  33. [33]

    Syndrone-multi-modal uav dataset for ur- ban scenarios

    Giulia Rizzoli, Francesco Barbato, Matteo Caligiuri, and Pietro Zanuttigh. Syndrone-multi-modal uav dataset for ur- ban scenarios. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 2210–2220,

  34. [34]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 4104–4113, 2016. 2

    2016

  35. [35]

    A comparison and evalua- tion of multi-view stereo reconstruction algorithms

    Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A comparison and evalua- tion of multi-view stereo reconstruction algorithms. In2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), pages 519–528. IEEE, 2006. 2

    2006

  36. [36]

    Flod: Integrating flexible level of detail into 3d gaussian splatting for customizable rendering, 2024

    Yunji Seo, Young Sun Choi, Hyun Seung Son, and Youngjung Uh. Flod: Integrating flexible level of detail into 3d gaussian splatting for customizable rendering, 2024. 3

    2024

  37. [37]

    A general albedo recovery approach for aerial photogrammetric images through inverse rendering.ISPRS Journal of Photogrammetry and Remote Sensing, 218:101–119, 2024

    Shuang Song and Rongjun Qin. A general albedo recovery approach for aerial photogrammetric images through inverse rendering.ISPRS Journal of Photogrammetry and Remote Sensing, 218:101–119, 2024. 2

    2024

  38. [38]

    Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs

    Haithem Turki, Deva Ramanan, and Mahadev Satya- narayanan. Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12922–12931, 2022. 2, 3

    2022

  39. [39]

    Dn-splatter: Depth and normal priors for gaussian splatting and meshing, 2024

    Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splatting and meshing, 2024. 3

    2024

  40. [40]

    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. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 2025. 3

    2025

  41. [41]

    Moge-2: Accurate monocular geometry with metric scale and sharp details, 2025

    Ruicheng Wang, Sicheng Xu, Yue Dong, Yu Deng, Jianfeng Xiang, Zelong Lv, Guangzhong Sun, Xin Tong, and Jiaolong Yang. Moge-2: Accurate monocular geometry with metric scale and sharp details, 2025. 5

    2025

  42. [42]

    Uavscenes: A multi-modal dataset for uavs

    Sijie Wang, Siqi Li, Yawei Zhang, Shangshu Yu, Shenghai Yuan, Rui She, Quanjiang Guo, JinXuan Zheng, Ong Kang Howe, Leonrich Chandra, et al. Uavscenes: A multi-modal dataset for uavs. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 28946–28958,

  43. [43]

    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. 2020. 3

    2020

  44. [44]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 6

    2004

  45. [45]

    Trun- cated signed distance function: experiments on voxel size

    Diana Werner, Ayoub Al-Hamadi, and Philipp Werner. Trun- cated signed distance function: experiments on voxel size. InImage Analysis and Recognition: 11th International Con- ference, ICIAR 2014, Vilamoura, Portugal, October 22-24, 2014, Proceedings, Part II 11, pages 357–364. Springer,

    2014

  46. [46]

    Surface re- construction from gaussian splatting via novel stereo views

    Yaniv Wolf, Amit Bracha, and Ron Kimmel. Surface re- construction from gaussian splatting via novel stereo views. arXiv preprint arXiv:2404.01810, 2024. 3

    Pith/arXiv arXiv 2024

  47. [47]

    Uavd4l: A large-scale dataset for uav 6-dof localization

    Rouwan Wu, Xiaoya Cheng, Juelin Zhu, Yuxiang Liu, Mao- jun Zhang, and Shen Yan. Uavd4l: A large-scale dataset for uav 6-dof localization. In2024 International Conference on 3D Vision (3DV), pages 1574–1583. IEEE, 2024. 3

    2024

  48. [48]

    En- vgs: Modeling view-dependent appearance with environ- ment gaussian

    Tao Xie, Xi Chen, Zhen Xu, Yiman Xie, Yudong Jin, Yu- jun Shen, Sida Peng, Hujun Bao, and Xiaowei Zhou. En- vgs: Modeling view-dependent appearance with environ- ment gaussian. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5742–5751, 2025. 2, 3

    2025

  49. [49]

    Gauu-scene v2: Assessing the reliability of image-based metrics with expansive lidar image dataset using 3dgs and nerf.arXiv preprint arXiv:2404.04880, 2024

    Butian Xiong, Nanjun Zheng, Junhua Liu, and Zhen Li. Gauu-scene v2: Assessing the reliability of image-based metrics with expansive lidar image dataset using 3dgs and nerf.arXiv preprint arXiv:2404.04880, 2024. 2, 3

    Pith/arXiv arXiv 2024

  50. [50]

    Gauu-scene v2: Assessing the reliability of image-based metrics with expansive lidar image dataset using 3dgs and nerf.CoRR, 2024

    Butian Xiong, Nanjun Zheng, Junhua Liu, and Zhen Li. Gauu-scene v2: Assessing the reliability of image-based metrics with expansive lidar image dataset using 3dgs and nerf.CoRR, 2024. 2

    2024

  51. [51]

    Crossloc: Scalable aerial localization as- sisted by multimodal synthetic data

    Qi Yan, Jianhao Zheng, Simon Reding, Shanci Li, and Ior- dan Doytchinov. Crossloc: Scalable aerial localization as- sisted by multimodal synthetic data. InProceedings of 10 the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17358–17368, 2022. 3

    2022

  52. [52]

    Spec-gaussian: Anisotropic view-dependent ap- pearance for 3d gaussian splatting, 2024

    Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xi- aoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xi- aogang Jin. Spec-gaussian: Anisotropic view-dependent ap- pearance for 3d gaussian splatting, 2024. 3

    2024

  53. [53]

    Reflective gaussian splatting.arXiv preprint arXiv:2412.19282, 2024

    Yuxuan Yao, Zixuan Zeng, Chun Gu, Xiatian Zhu, and Li Zhang. Reflective gaussian splatting.arXiv preprint arXiv:2412.19282, 2024. 2, 3, 5, 6

    Pith/arXiv arXiv 2024

  54. [54]

    Stablenormal: Reducing diffusion variance for stable and sharp normal.ACM Transactions on Graphics (TOG), 2024

    Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal.ACM Transactions on Graphics (TOG), 2024. 5

    2024

  55. [55]

    3d gaussian splat- ting with deferred reflection

    Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splat- ting with deferred reflection. InACM SIGGRAPH 2024 Con- ference Papers, pages 1–10, 2024. 3

    2024

  56. [56]

    Absgs: Recovering fine details for 3d gaussian splat- ting, 2024

    Zongxin Ye, Wenyu Li, Sidun Liu, Peng Qiao, and Yong Dou. Absgs: Recovering fine details for 3d gaussian splat- ting, 2024. 5, 8

    2024

  57. [57]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 19447–19456,

  58. [58]

    Gaussian opacity fields: Efficient high-quality compact surface recon- struction in unbounded scenes.arXiv:2404.10772, 2024

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient high-quality compact surface recon- struction in unbounded scenes.arXiv:2404.10772, 2024. 3, 5, 7

    Pith/arXiv arXiv 2024

  59. [59]

    Dogaussian: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaus- sian consensus

    Gim Hee Lee Yu Chen. Dogaussian: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaus- sian consensus. InarXiv, 2024. 3

    2024

  60. [60]

    Rade-gs: Rasterizing depth in gaussian splatting.arXiv preprint arXiv:2406.01467, 2024

    Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting.arXiv preprint arXiv:2406.01467, 2024. 3

    Pith/arXiv arXiv 2024

  61. [61]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    2018

  62. [62]

    Ref-gs: Directional factorization for 2d gaussian splatting

    Youjia Zhang, Anpei Chen, Yumin Wan, Zikai Song, Jun- qing Yu, Yawei Luo, and Wei Yang. Ref-gs: Directional factorization for 2d gaussian splatting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26483–26492, 2025. 2, 3, 5, 6

    2025

  63. [63]

    Aerialgo: Walking-through city view generation from aerial perspectives.arXiv preprint arXiv:2412.00157, 2024

    Fuqiang Zhao, Yijing Guo, Siyuan Yang, Xi Chen, Luo Wang, Lan Xu, Yingliang Zhang, Yujiao Shi, and Jingyi Yu. Aerialgo: Walking-through city view generation from aerial perspectives.arXiv preprint arXiv:2412.00157, 2024. 2, 3

    Pith/arXiv arXiv 2024

  64. [64]

    Tihui Zhou, Linbing Lv, Jinhu Liu, and J Wan. Application of uav oblique photography in real scene 3d modeling.The International Archives of the Photogrammetry, Remote Sens- ing and Spatial Information Sciences, 43:413–418, 2021. 2 11 Beyond a Single Light: A Large-Scale Aerial Dataset for Urban Scene Reconstruction Under Varying Illumination Supplementar...

    2021

  65. [65]

    Equipment Details Table 5

    Dataset Details 6.1. Equipment Details Table 5. Specifications of the UA V platform, oblique camera, and LiDAR payload used in our data acquisition. Device Type Key parameters DJI M350 RTKUA V platform

  66. [66]

    Max flight time:55 min (no payload)

  67. [67]

    CHCNA V C30Oblique camera

    Sensing: 6-direction. CHCNA V C30Oblique camera

  68. [68]

    Total resolution:130 MP (26 MP×5)

  69. [69]

    Image size:6252×4168(3:2)

  70. [70]

    Lenses:1×∼nadir90◦+ 4×∼oblique45◦

  71. [71]

    Focal lengths:25 mm / 35 mm

  72. [72]

    Minimum capture interval:0.8 s

  73. [73]

    DJI L2LiDAR

    Dimensions:110×108×85mm. DJI L2LiDAR

  74. [74]

    LiDAR range:450 m (50% reflectivity)

  75. [75]

    Point rate:1.2M pts/s (multi-return)

  76. [76]

    Accuracy:5 cm horizontal / 4 cm vertical

  77. [77]

    Ranging accuracy:2 cm @ 150 m

  78. [78]

    FOV(non-repetitive):70◦×75◦

  79. [79]

    RGB sensor:4/3” CMOS, 20 MP

  80. [80]

    It is an industrial-rank UA V with a maximum flight time of about 55 minutes with- out payload

    Weight:905 g; UA V Platform (DJI M350 RTK).We use a DJI Matrice 350 RTK as the carrier platform. It is an industrial-rank UA V with a maximum flight time of about 55 minutes with- out payload. The built-in RTK module provides centimeter- level positioning when used with network RTK or a base station. These properties enable us to replicate nearly iden- ti...

Showing first 80 references.