SkyLume | Aerial 3D Reconstruction Under Varying Illumination

Overview

A benchmark for illumination-robust city-scale reconstruction.

Existing aerial reconstruction pipelines often bake observed lighting into textures, radiance fields, or Gaussians. SkyLume turns that nuisance into an explicit benchmark: each scene is revisited along the same RTK-guided flight path at three times of day, making cross-time robustness measurable rather than incidental.

The release is designed around geometry, appearance, and reproducibility. It provides five synchronized views per capture, unified 6-DoF poses, LiDAR-guided meshes, per-frame depth and normals, and solar geometry annotations for future de-shadowing, relighting, and inverse-rendering studies.

Repeated illumination

Morning, noon, and afternoon captures isolate lighting changes while keeping scene structure and viewpoints comparable.

Survey-grade geometry

LiDAR scans regularize weak-texture, shadowed, reflective, and water-adjacent regions during mesh construction.

Cross-time metrics

The Temporal Consistency Coefficient evaluates whether albedo and geometry remain stable across illumination periods.

Dataset

From repeatable UAV flights to aligned, multi-modal supervision.

Aligned camera trajectories from all three illumination periods — Aligned trajectories from all three illumination periods.

Single-period camera trajectory for Period 0 after alignment — Single-period camera trajectory split for Period 0.

Single-period camera trajectory for Period 1 after alignment — Single-period camera trajectory split for Period 1.

Single-period camera trajectory for Period 2 after alignment — Single-period camera trajectory split for Period 2.

Capture protocol

Each region is flown in three daily slots along the same trajectory, with 80% forward overlap, 60% side overlap, about 120 m flight height, and 1 Hz camera triggering.

Sensor stack

The C30 camera records one nadir view and four oblique views at 26 MP per lens. The DJI L2 LiDAR provides centimeter-level metric support for geometry evaluation.

Released modalities

SkyLume exports COLMAP-style SfM packages, per-period poses, mesh depth, LiDAR depth, mesh normals, post-processed meshes, and solar angles.

A preview dataset is available on OneDrive. The full dataset will be uploaded to Hugging Face.

Six representative SkyLume scenes and dataset statistics — Representative medium-scale scenes and dataset statistics: 10 scenes, 7.21 km², three illumination periods, and more than 100,000 UAV images.

Per-scene statistics

Scale	Scene	Total images	Flight height	Density	Illumination setup	Water	Glass
Small	Gym	6,185	120.366 m	Low	Sunlight / Sunlight / Overcast	No	No
Small	Staff Residence	7,920	130.018 m	Medium	Sunlight / Partly cloudy / Overcast	Yes	No
Small	iPark	5,355	115.241 m	Medium	Sunlight / Partly cloudy / Overcast	Yes	Yes
Medium	Tec School	7,185	108.481 m	Low	Sunlight / Sunlight / Sunlight	Yes	No
Medium	Buildings	10,455	129.891 m	High	Sunlight / Sunlight / Overcast	No	Yes
Medium	High School	10,065	108.015 m	Medium	Sunlight / Partly cloudy / Sunlight	Yes	No
Medium	Main Campus	10,410	130.280 m	Medium	Sunlight / Partly cloudy / Sunlight	No	No
Large	Estate	18,630	109.024 m	High	Sunlight / Partly cloudy / Sunlight	No	No
Large	Town	12,435	149.260 m	High	Sunlight / Partly cloudy / Overcast	Yes	Yes
Large	Med School	20,700	119.250 m	Medium	Sunlight / Sunlight / Overcast	Yes	No

LiDAR-guided ground-truth meshes for representative SkyLume scenes — Post-processed LiDAR-guided meshes provide ground truth for representative medium-scale scenes.

Partly cloudy illumination causing inconsistent local lighting — Partly cloudy flights create local lighting variation, testing whether methods overfit shadows and exposure drift.

Positioning

SkyLume covers the axes missing from prior aerial datasets.

Dataset	Real-world	LiDAR	Camera type	Light variation	Depth / normal	Resolution
ISPRS Benchmark	Yes	Terrestrial	Oblique	No	No	6000 x 4000
UrbanScene3D	Partial	Yes	Oblique	No	No	5490 x 3651
GauU-Scene	Yes	Yes	Oblique + Nadir	No	No	5468 x 3636
MatrixCity	No	From depth	Oblique + Nadir	Yes	Depth + normal	1920 x 1080
SkyLume	Yes	Yes	Oblique + Nadir	Yes	Depth + normal	6252 x 4168

Benchmark

Evaluation tracks for appearance, geometry, and novel views.

TCC-Albedo

Inverse-rendering methods are evaluated by rendering albedo from matched viewpoints across three time slots.

Best reported mean overall: 0.775

TCC-Geometry

Meshes are compared to ground truth and against one another across illumination periods using F-1 consistency.

Best F-1 at 0.5 m: 0.719

Novel View Synthesis

NVS quality is reported with PSNR, SSIM, and LPIPS on six UAV scenes under challenging sunlit captures.

Abs-GS is strongest overall

Temporal Consistency Coefficient evaluation protocol — TCC evaluates albedo and geometry by fixing viewpoints across periods and scoring cross-time consistency.

Albedo and TCC visualization across two illumination periods — Inverse-rendering baselines retain time-dependent shadows and exposure changes in estimated albedo.

Novel view synthesis qualitative comparisons on SkyLume scenes — NVS baselines vary sharply in shadow-dominated and weak-texture aerial regions.

Cross-time rendering and mesh comparison under different periods — Strong lighting shifts introduce shadow-as-geometry artifacts and cross-view color inconsistencies.

Representative benchmark numbers

Track	Method	Metric	Score	Interpretation
TCC-Albedo	GS-IR	Mean overall	0.721	Cross-time albedo consistency
TCC-Albedo	Ref-Gaussian	Mean overall	0.658	Cross-time albedo consistency
TCC-Albedo	Ref-GS	Mean overall	0.775	Best reported mean in the draft
TCC-Geometry	2DGS	F-1 at 0.5 m	0.675	Average pairwise consistency
TCC-Geometry	CityGaussianV2	F-1 at 0.5 m	0.719	Best reported geometry consistency

Citation

Cite SkyLume

@misc{li2026singlelightlargescaleaerial,
        title={Beyond a Single Light: A Large-Scale Aerial Dataset for Urban Scene Reconstruction Under Varying Illumination}, 
        author={Zhuoxiao Li and Wenzong Ma and Taoyu Wu and Jinjing Zhu and Shuai Zhang and Jing OU and Tongyan Hua and Yinrui Ren and Rongjun Qin and Hui Xiong and Wufan Zhao},
        year={2026},
        eprint={2512.14200},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2512.14200}, 
}