--- language: - en tags: - relighting - multi-view - 3d - objaverse - hdr - rendering - depth-maps task_categories: - image-to-image - depth-estimation license: cc-by-4.0 size_categories: - 10M

Teaser

vLAR Group | SIGGRAPH Asia 2026

A large-scale, high-quality dataset for multi-view relighting.

--- ## 📖 Dataset Summary The **Laval Objaverse Dataset** is a comprehensive dataset designed for multi-view relighting and novel view synthesis tasks. It combines high-quality 3D assets from Objaverse with realistic, diverse illumination conditions from the Laval Indoor and Outdoor HDR datasets. Each render includes synchronized multi-view images, depth maps, and complete lighting metadata. ## 🏗️ Dataset Construction We structured our dataset through a rigorous four-step pipeline: 1. **🎯 Object Filtering:** We source base meshes from the [Objaverse](https://objaverse.allenai.org/) dataset. To ensure high visual fidelity, we exclude meshes with poor geometry or materials by adopting the strict object selection criteria from the relitObjaverse dataset [[1]](#references). 2. **💡 Lighting Selection:** We employ open-source illumination datasets, specifically the *Laval Indoor HDR Dataset* [[2]](#references) and the *Laval Outdoor HDR Dataset* [[3]](#references). Each illumination map is uniformly rotated 16 times to maximize lighting diversity. For every object, we sample 16 environment maps with a balanced 1:1 ratio of indoor and outdoor lighting conditions. 3. **📷 Rendering Protocol:** All objects are normalized to unit size (largest bounding box dimension = 1) and placed at the origin. Cameras are uniformly distributed on a sphere at distances of 1.8 or 2.2 units from the center, oriented toward the scene center. 4. **🛡️ Dataset Partitioning:** The dataset is split into **training, validation, and test sets** with strict disjointness: no camera poses, illumination conditions, or 3D objects overlap between splits. The division information are under **`info/laval`**, **`info/objaverse`** and **`info/view`** for illuminations, objects and viewpoints respecitively. ## 📂 Dataset Structure ```text LavalObjaverseDataset/ └── info/ │ ├── laval/ │ │ └── full_testing_lighting.json │ │ └── full_training_lighting.json │ │ └── full_validation_lighting.json │ │ │ ├── objaverse/ │ │ └── all_objects.json │ │ └── full_{split}_objects.json # split = testing, training, validation │ │ │ ├── view/ │ └── all_views.json │ └── full_{split}_views.json # split = testing, training, validation │ └── pairs/ │ ├── 1_to_1_mapping_pairs.json │ ├── 16_to_16_mapping_pairs.json │ ├── 32_to_32_mapping_pairs.json │ ├── olatverse.json │ └── ... │ └── rendered/ ├── training/ │ └── subset_{subset_uid}/ │ └── {object_uid_0}.tar.gz │ │ ├── info.json │ │ ├── V18-Indoor_9C4A5417-f5d395fb63_6_image.png │ │ ├── V18-Outdoor_9C4A0573-f5d395fb63_6_image.png │ │ ├── V18_depth_0001.exr │ │ ├── V23-Indoor_9C4A5417-f5d395fb63_6_image.png │ │ └── ... │ └── ... │ ├── testing/ │ ├── {object_uid_1}.tar.gz │ ├── {object_uid_2}.tar.gz │ └── ... │ └── validation/ ├── {object_uid_3}.tar.gz ├── {object_uid_4}.tar.gz └── ... ``` ### Inside Each Object Folder / Unzipped Archive: - **`info.json`**: Metadata containing object information, rendering parameters, camera poses, and lighting conditions. - **`Vxxx-yyy_image.png`**: Rendered RGB images where: - `Vxxx` indicates the camera pose/viewpoint (e.g., `V18`, `V23`). - `yyy` is the lighting map name (e.g., `Indoor_9C4A5417-f5d395fb63_6` or `Outdoor_9C4A0573-f5d395fb63_6`). - **`Vxxx_depth_*.exr`**: Corresponding ground-truth depth maps in OpenEXR format for each viewpoint. ### File Naming Convention Example: ```text V18-Indoor_9C4A5417-f5d395fb63_6_image.png │ │ │ │ │ │ │ └─ Rotating number (0,1,...,15) │ │ │ │ │ └─ Lighting map identifier(9C4A5417-f5d395fb63, etc.) │ └─ Lighting type (Indoor/Outdoor) └─ Viewpoint ID (V18, V23, etc.) ``` ## 📚 References If you use this dataset in your research, please cite our work along with the prior work we built upon: ```bibtex @misc{wang2026relightformerfeedforwardgenerativetransformer, title={RelightFormer: Feed-forward Generative Transformer for Multiview Object Relighting}, author={Hejun Wang and Jinxi Li and Junwei Jiang and Shiwei Mao and Hu Cheng and Shouwang Huang and Bo Yang}, year={2026}, eprint={2609.07414}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2609.07414}, } @article{Jin2024, author = {Jin, Haian and Li, Yuan and Luan, Fujun and Xiangli, Yuanbo and Bi, Sai and Zhang, Kai and Xu, Zexiang and Sun, Jin and Snavely, Noah}, journal = {NeurIPS}, title = {{Neural Gaffer: Relighting Any Object via Diffusion}}, year = {2024} } @article{laval_indoor, author = {Marc-Andr\'{e} Gardner and Kalyan Sunkavalli and Ersin Yumer and Xiaohui Shen and Emiliano Gambaretto and Christian Gagn\'{e} and Jean-Fran\c{c}ois Lalonde}, title = {Learning to Predict Indoor Illumination from a Single Image}, journal = {SIGGRAPH Asia}, year = {2017} } @INPROCEEDINGS{laval_outdoor, author={Hold-Geoffroy, Yannick and Athawale, Akshaya and Lalonde, Jean-François}, booktitle={CVPR}, title={Deep Sky Modeling for Single Image Outdoor Lighting Estimation}, year={2019} } ``` ## 📮 Contact For questions about the dataset, please contact: - {hejun.wang, jinxi.li}@connect.polyu.hk - bo.yang@polyu.edu.hk --- *For rendering code, models, and further details, please visit our [GitHub Repository](https://github.com/vLAR-group/RelightFormer).*