--- license: cc-by-nc-4.0 pretty_name: OmniRooms task_categories: - depth-estimation - image-to-3d tags: - panorama - depth - synthetic - indoor-scene - 3d - vla size_categories: - 100K UniSHARP:
Universal Sharp Monocular View Synthesis

Meixi Song¹ · Dizhe Zhang^1,* · Hao Ren¹ · Ruiyang Zhang¹ · Bo Du² · Ming-Hsuan Yang³ · Lu Qi^1,2,*
¹Insta360 Research · ²Wuhan University · ³University of California, Merced

UniSHARP extends SHARP-style photorealistic monocular view synthesis to universal camera systems. Given a single image from a perspective, wide-FoV, fisheye, or panoramic camera, UniSHARP predicts a 3D Gaussian representation and renders high-quality novel views. # 🏠 OmniRooms OmniRooms is a large-scale synthetic indoor dataset for panoramic 3D perception, depth estimation, and monocular view synthesis. It is designed for research on 3D reconstruction and 3D Gaussian Splatting from universal camera inputs, especially equirectangular panoramas and wide-FoV fisheye images. The dataset contains 16 large indoor scenes and each scene contains multiple rooms. In total, the current release contains 271k equirectangular RGB panoramas and paired depth maps, plus 11,880 derived fisheye images for OmniRooms-Wide. # 🖼️ Scene Preview ## 🏘️ 16 Scenes

## 🎯 One Group

One Group

## ⚙️ Data Generation OmniRooms is collected in AirSim environments. Anchor points are sampled on a 0.5 m voxel grid. For each anchor point, we render one central camera and 29 cameras randomly sampled within a local axis-aligned 30 cm cube centered on the source camera. To isolate translation-induced synthesis controlled, all cameras share a fixed orientation in the panorama release. Because this is a panorama, you can manually create rotations through code. OmniRooms-Wide is generated by projecting OmniRooms panoramas into 130-degree equidistant fisheye views. Each panorama frame is rendered as a 1024 x 2048 equirectangular RGB image. Depth maps are provided at the same resolution. ## 📁 Directory Structure The dataset root is organized by scene: ```text OmniRooms/ ├── / │ ├── Original/ │ │ ├── Panoramagram0.jpg │ │ ├── Panoramagram1.jpg │ │ └── ... │ └── Depth/ │ ├── Depth0.h5 │ ├── Depth1.h5 │ └── ... └── 30cm (pose)/ ├── AI_vol3_01.csv └── ... ``` ### RGB panoramas RGB panoramas are stored as JPEG files: ```text /Original/Panoramagram{index}.jpg ``` Each RGB panorama is a 2048 x 1024 equirectangular image in standard RGB color space. ### Depth maps Depth files are stored as HDF5 files: ```text /Depth/Depth{index}.h5 ``` Each HDF5 depth file contains: - `depth`: float32 array with shape `(1024, 2048)`. - `alpha`: float32 array with shape `(1024, 2048)`. ### Camera positions Camera positions are provided as CSV files: ```text 30cm/.csv ``` The CSV files use columns: ```text ,X,Y,Z ``` ## 💻 Loading Example The following example loads one RGB panorama and its paired depth map: ```python from pathlib import Path import h5py from PIL import Image root = Path("/path/to/OmniRooms") scene = "AI_vol3_01" frame_idx = 0 rgb_path = root / scene / "Original" / f"Panoramagram{frame_idx}.jpg" depth_path = root / scene / "Depth" / f"Depth{frame_idx}.h5" rgb = Image.open(rgb_path).convert("RGB") with h5py.File(depth_path, "r") as f: depth = f["depth"][:] print(rgb.size) # (2048, 1024) print(depth.shape) # (1024, 2048) ``` ## 📜 License This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). You may use the dataset for non-commercial research and education. Commercial use is not permitted without additional permission. ## 📚 Citation If you use OmniRooms or OmniRooms-Wide in your research, please cite the corresponding UniSHARP project: ```bibtex @article{song2026unisharp, title={UniSHARP: Universal Sharp Monocular View Synthesis}, author={Song, Meixi and Zhang, Dizhe and Ren, Hao and Zhang, Ruiyang and Du, Bo and Yang, Ming-Hsuan and Qi, Lu}, journal={arXiv}, year={2026} } ```