--- license: cc-by-4.0 dataset_info: - config_name: default features: - name: frame_id dtype: int32 - name: orig_frame dtype: string - name: image dtype: image - name: image_raw dtype: binary - name: albedo dtype: image - name: albedo_raw dtype: binary - name: depth dtype: image - name: depth_raw dtype: binary - name: normal dtype: image - name: normal_raw dtype: binary - name: mask dtype: image - name: camera dtype: string - name: sky_raw dtype: binary - name: model dtype: string - name: scene dtype: string - name: date dtype: string - name: lighting dtype: string splits: - name: train num_bytes: 48468861351 num_examples: 5664 - name: test num_bytes: 9440057784 num_examples: 520 - name: train_selected num_bytes: 21972048865 num_examples: 2487 download_size: 79310118605 dataset_size: 79880968000 - config_name: models features: - name: model_name dtype: string - name: scene dtype: string - name: date dtype: string - name: lighting dtype: string - name: model_raw dtype: binary splits: - name: train num_bytes: 7294616740 num_examples: 27 download_size: 3751431547 dataset_size: 7294616740 configs: - config_name: default data_files: - split: train path: data/train-* - split: test path: data/test-* - split: train_selected path: data/train_selected-* - config_name: models data_files: - split: train path: models/train-* --- # Olbedo: An Albedo and Shading Aerial Dataset for Large-Scale Outdoor Environments [Shuang Song](https://openreview.net/profile?id=~Shuang_Song10)¹‡, [Debao Huang](https://debaohuang.github.io/)¹‡, [Deyan Deng](https://openreview.net/profile?id=~Deyan_Deng1)¹, [Haolin Xiong](https://openreview.net/profile?id=~Haolin_Xiong1)², [Yang Tang](https://openreview.net/profile?id=~Yang_Tang4)¹, [Yajie Zhao](https://openreview.net/profile?id=~Yajie_Zhao1)², [Rongjun Qin](https://openreview.net/profile?id=~Rongjun_Qin1)¹* ¹ The Ohio State University · ² University of Southern California ‡ Equal contribution · * Corresponding author This repository contains the Olbedo dataset, containing RGB images, albedo, depth, surface normals, camera parameters, sky HDR maps, and 3D scene models. **HuggingFace:** [GDAOSU/Olbedo](https://huggingface.co/datasets/GDAOSU/Olbedo) ## Dataset Structure ### Main config (default) Three splits: `train` (5,664 samples), `train_selected` (2,487 samples), `test` (520 samples). | Column | Type | Description | |--------|------|-------------| | `frame_id` | int | Unique frame identifier | | `orig_frame` | string | Original frame number | | `image` | Image | sRGB preview (PNG, for visualization) | | `image_raw` | binary | Original EXR file (Linear RGB, for training) | | `albedo` | Image | sRGB preview (PNG, for visualization) | | `albedo_raw` | binary | Original EXR file (Linear RGB, for training) | | `depth` | Image | Colormap preview (PNG, INFERNO colormap) | | `depth_raw` | binary | Original EXR file (float, for training) | | `normal` | Image | RGB preview (PNG, remapped from [-1,1] to [0,1]) | | `normal_raw` | binary | Original EXR file (float RGB, for training) | | `mask` | Image | Segmentation mask (PNG) | | `camera` | string | Camera parameters (JSON string) | | `sky_raw` | binary | Sky HDR map (.hdr), `None` if not available | | `model` | string | GLB model filename (e.g. `scene_date_lighting.glb`) | | `scene` | string | Scene name | | `date` | string | Capture date (YYYYMMDD) | | `lighting` | string | Lighting condition (sunrise/sunset/overcast) | The `test` split has `None` for depth, normal, camera, sky, and model columns. ### Models config A separate config containing 27 3D scene models in GLB format. | Column | Type | Description | |--------|------|-------------| | `model_name` | string | Model identifier (e.g. `osu_coe_corridors_20220720_sunrise`) | | `scene` | string | Scene name | | `date` | string | Capture date | | `lighting` | string | Lighting condition | | `model_raw` | binary | GLB file (binary) | ## Usage ### Load the dataset ```python from datasets import load_dataset # Load a split (downloads data) ds = load_dataset("GDAOSU/Olbedo", split="train") # Or use streaming to avoid downloading everything ds = load_dataset("GDAOSU/Olbedo", split="train", streaming=True) row = next(iter(ds)) ``` ### Recover raw EXR files (image, albedo, depth, normal) ```python ds = load_dataset("GDAOSU/Olbedo", split="train", streaming=True) for row in ds: fid = row['frame_id'] # Save image EXR with open(f"{fid:04d}_im.exr", "wb") as f: f.write(row['image_raw']) # Save albedo EXR with open(f"{fid:04d}_albedo.exr", "wb") as f: f.write(row['albedo_raw']) # Save depth EXR if row['depth_raw'] is not None: with open(f"{fid:04d}_depth.exr", "wb") as f: f.write(row['depth_raw']) # Save normal EXR if row['normal_raw'] is not None: with open(f"{fid:04d}_normal.exr", "wb") as f: f.write(row['normal_raw']) ``` ### Recover camera JSON ```python import json row = next(iter(ds)) if row['camera'] is not None: camera = json.loads(row['camera']) print(camera['focal'], camera['cx'], camera['cy']) # intrinsics print(camera['X'], camera['Y'], camera['Z']) # translation # Save to file fid = row['frame_id'] with open(f"{fid:04d}_camera.json", "w") as f: f.write(row['camera']) ``` ### Recover sky HDR maps Only some frames have sky HDR maps (frames 3801+). ```python for row in ds: if row['sky_raw'] is not None: fid = row['frame_id'] with open(f"{fid:04d}_sky.hdr", "wb") as f: f.write(row['sky_raw']) ``` ### Recover 3D models (GLB) Models are stored in a separate config. Each model corresponds to a unique scene/date/lighting combination. ```python models = load_dataset("GDAOSU/Olbedo", "models", split="train", streaming=True) for row in models: name = row['model_name'] with open(f"{name}.glb", "wb") as f: f.write(row['model_raw']) print(f"Saved {name}.glb ({len(row['model_raw']) / 1e6:.0f} MB)") ``` ### Link frames to their 3D model Each frame's `model` field contains the GLB filename. To find which model a frame belongs to: ```python row = next(iter(ds)) print(row['model']) # e.g. "osu_coe_corridors_20220720_sunrise.glb" ``` ### Recover all data for a single frame ```python import json ds = load_dataset("GDAOSU/Olbedo", split="train", streaming=True) row = next(iter(ds)) fid = row['frame_id'] prefix = f"{fid:04d}" # Save all files with open(f"{prefix}_im.exr", "wb") as f: f.write(row['image_raw']) with open(f"{prefix}_albedo.exr", "wb") as f: f.write(row['albedo_raw']) if row['depth_raw']: with open(f"{prefix}_depth.exr", "wb") as f: f.write(row['depth_raw']) if row['normal_raw']: with open(f"{prefix}_normal.exr", "wb") as f: f.write(row['normal_raw']) if row['camera']: with open(f"{prefix}_camera.json", "w") as f: f.write(row['camera']) if row['sky_raw']: with open(f"{prefix}_sky.hdr", "wb") as f: f.write(row['sky_raw']) # Save mask from preview Image if row['mask'] is not None: row['mask'].save(f"{prefix}_mask.png") print(f"Frame {fid}: scene={row['scene']}, date={row['date']}, lighting={row['lighting']}, model={row['model']}") ``` ## Scenes The dataset covers 4 locations with multiple captures under different dates and lighting conditions: - `goodale_park` - `osu_coe_corridors` - `osu_residential_area` - `schottenstein_center` ## File Formats - **Image/Albedo EXR**: Linear RGB, float16/float32. Apply sRGB transfer function for display. - **Depth EXR**: Single-channel float (channel name `I`). Use OpenEXR library to read. - **Normal EXR**: RGB float in [-1, 1] range. Remap to [0, 1] for visualization. - **Sky HDR**: Radiance HDR format (.hdr). - **3D Models**: glTF Binary (.glb). - **Camera JSON**: Contains intrinsics (focal, cx, cy, distortion), extrinsics (rotation matrix, translation), and metadata (GPS, sun position).