Intrinsic image decomposition (IID) of outdoor scenes is crucial for relighting, editing, and understanding large-scale environments, but progress has been limited by the lack of real-world datasets with reliable albedo and shading supervision. We introduce Olbedo, a large-scale aerial dataset for outdoor albedo--shading decomposition in the wild. Olbedo contains 5,664 UAV images captured across four landscape types, multiple years, and diverse illumination conditions. Each view is accompanied by multi-view consistent albedo and shading maps, metric depth, surface normals, sun and sky shading components, camera poses, and, for recent flights, measured HDR sky domes. These annotations are derived from an inverse-rendering refinement pipeline over multi-view stereo reconstructions and calibrated sky illumination, together with per-pixel confidence masks. We demonstrate that Olbedo enables state-of-the-art diffusion-based IID models, originally trained on synthetic indoor data, to generalize to real outdoor imagery: fine-tuning on Olbedo significantly improves single-view outdoor albedo prediction on the MatrixCity benchmark. We further illustrate applications of Olbedo-trained models to multi-view consistent relighting of 3D assets, material editing, and scene change analysis for urban digital twins. We release the dataset, baseline models, and an evaluation protocol to support future research in outdoor intrinsic decomposition and illumination-aware aerial vision.
Figures & tables
Figure 2: Landscapes of the Olbedo dataset. (Credit: Google)
Figure 3: Sky radiance acquisition and representation. Left: exposure-bracketed raw sky images. Right: merged HDR sky dome in equirectangular format.
Figure 4: Statistics of our multi-temporal albedo dataset. (a) Distribution of images per year. (b) Distribution of images per site. (c) Distribution of images per lighting condition.
Figure 5: Our image decomposition framework.
Figure 6: Qualitative comparison of fine-tuned and pretrained models for outdoor albedo prediction on the MatrixCity dataset. The first three rows show original image and albedo, and the last two rows show per-pixel absolute error maps for the fine-tunable networks only, where warmer colors indicate larger error and cooler colors indicate smaller error. CD = Colorful Diffuse [ 9 ] , IA = IntrinsicAnything [ 10 ] , IID = Intrinsic Image Diffusion [ 20 ] , Marigold = Marigold-IID-Appearance [ 19 ] , RGBX = RGB ↔ X [ 44 ] , and NIID = NIID-Net [ 27 ] .
Model
PSNR ↑
SSIM ↑
LPIPS ↓
Pretrained IA
13.422
0.478
0.910
Pretrained CD
14.905
0.545
0.554
Pretrained NIID
12.782
0.549
0.793
Fine-tuned NIID
16.152
0.594
0.769
Pretrained IID
15.594
0.493
0.554
Fine-tuned IID
17.249
0.531
0.485
Table 1: Quantitative comparison of fine-tuned and pretrained models for outdoor albedo prediction on the MatrixCity dataset. IA = IntrinsicAnything [ 10 ] , CD = Colorful Diffuse [ 9 ] , NIID = NIID-Net [ 27 ] , IID = Intrinsic Image Diffusion [ 20 ] , Marigold = Marigold-IID-Appearance [ 19 ] , and RGBX = RGB ↔ X [ 44 ] .
Figure 7: Predicted albedo on a BlendedMVS scene from the fine-tuned RGB ↔ X model. Across viewpoints, baked-in cast shadows and shading gradients are largely removed while roofs, walls, and roads retain a consistent appearance, making the results suitable for multi-view retexturing.
Figure 8: Segmentation with SAM on RGB and albedo using the same point of interest. From left to right, the columns show the RGB image, predicted albedo, SAM mask from RGB, SAM mask from albedo, and the two overlays. Albedo suppresses shadow-induced edges and yields more coherent masks, especially when cast shadows interrupt object continuity.
Figure 9: Material editing comparison on an aerial scene. The target colors are (RGB: 190, 115, 92), (RGB: 40, 0, 0), respectively. Alpha blending flattens the original shading, and AI editing baselines often alter illumination or wash out roof details, whereas our intrinsic-decomposition workflow preserves both the original lighting and local texture.
Figure 10: Lighting-invariant scene change analysis with RGB differencing and albedo differencing. The left three columns show the RGB reference, RGB source, and naive RGB difference; the right three columns show the corresponding albedo pair and albedo-based difference. Albedo suppresses false positives caused by shadows and illumination changes while preserving true scene changes such as parked vehicles.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1 : Comparison between in-camera JPEG (a) and RAW DNG (b) from the same scene. The JPEG image exhibits visible artifacts such as color clipping in highlights, loss of shadow details, and posterization effects due to aggressive compression and tone mapping. In contrast, the RAW DNG preserves a wider dynamic range and more accurate color representation, which are essential for reliable albedo–shading decomposition.
Figure A.2 : Additional examples from the Olbedo dataset. Rows from top to bottom: arena, office, residential, and park scenes.
Figure C.1 : Representative failure cases in pseudo-ground-truth generation. The top row shows the raw inputs and the bottom row shows the corresponding recovered albedo. Columns (a)–(e) illustrate geometry holes, sharp shadow-boundary noise, imperfect tree geometry, glass reflection, and glass transmission, respectively.
Figure D.1 : Example of the binary confidence mask used during fine-tuning. The mask is constructed by combining geometry boundaries and shadow-projection boundaries. Blue regions indicate high-confidence supervision, while dimmer regions are excluded from training.
Figure E.1 : Retextured models using recovered albedo images from the fine-tuned RGB ↔ X.
Figure E.2 : Comparison between RGB-textured models and albedo-retextured models. Shading is largely removed while texture details are well preserved.
Figure F.1 : Visual comparisons between the RGB textures (left column of each pair) and our recovered albedo textures (right column of each pair) under novel lighting conditions. The scenes encompass diverse architectural styles, materials, and environmental conditions. Original textures contain baked-in lighting such as cast shadows, shading gradients, and environment-induced color shifts, whereas our albedo textures remove these illumination effects and retain only the intrinsic diffuse reflectance. These results highlight the stability and consistency of our method across varied outdoor environments.
Figure G.1 : Additional segmentation results using SAM on RGB and albedo images. Each group of six images shows, from left to right: the RGB image, the corresponding albedo, the SAM mask predicted on RGB, the SAM mask predicted on albedo, and the two masks overlaid on the RGB image. Across diverse scenes, masks obtained from albedo remain more stable in shadowed regions and better preserve object boundaries than those obtained from RGB.
Recent advances in Intrinsic Image Decomposition (IID) have increasingly relied on generative models. However, progress remains limited by three key challenges: (a) insufficient physical consistency, (b) high computational cost at inference time, and (c) limited generalization capabilities. In this work, we show that latent bridge matching (LBM) effectively addresses these limitations for albedo estimation. We introduce a novel LBM-based architecture that enforces physical consistency through a pixel reconstruction loss, benefits from the inherent efficiency of LBM low-cost inference, and improves generalization across diverse datasets by incorporating a shading conditioning. In this extended version, we additionally show that conditioning the shading estimator itself on the predicted albedo further improves reconstruction fidelity, and we benchmark our best model against stateof-the-art IID methods across five real and synthetic datasets.
Carme Corbi, David Serrano-Lozano, Javier Vazquez-Corral +1
Universitat Aut`onoma de Barcelona and Computer Vision Center
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.
Zhuoxiao Li, Wenzong Ma, Taoyu Wu +8
HKUST(GZ) · University of Liverpool · The Ohio State University
We introduce SeasonScapes framework and a the SeasonScapes dataset: Swiss Sparse-view Mountain Scenes with Seasonal Changes that covers over 50 km x 60 km, composed of more than 85,000 webcam images captured from 32 different locations across 13 timestamps throughout a full year. By projecting these timestamp-specific images onto a 3D mesh, we construct seasonal 3D landscapes that reflect natural appearance changes over time. To address occlusions and missing data, we leverage conditional diffusion models for image-guided inpainting directly on the mesh. The resulting completed meshes can be further relighted using standard physically-based renderer.
Timo Kleger, Qi Ma, Deheng Zhang +2
2ETH Zurich · 1INSAIT, Sofia University “St. Kliment Ohridski”