Datasets:
DarkOcc · Real-world indoor RGB-D benchmark
10 scenes · 19 recording sequences · 8,208 paired RGB and depth frames
How much indoor geometry can a monocular occupancy model recover when the lights are dim? We recorded indoor rooms under normal and low light to evaluate this question on real camera inputs. The RGB-D portion of the benchmark is available now by request. We will release the complete dataset upon acceptance.
From a room recording to an occupancy reference
- Capture and align. Record RGB-D sequences and estimate camera poses to align observations across views.
- Build a room reference. Fuse depth into a room reconstruction, combine semantic image masks across views, and check the result against the source observations.
- Sample each frame. Use the camera pose to sample the room reference on a local voxel grid. Voxels with insufficient evidence remain unknown during evaluation.
The figure shows how we constructed the evaluation labels. The files in this release contain the paired RGB and depth observations.
Available recordings
| Illumination | Recording sequences | RGB-D pairs |
|---|---|---|
Normal light (bright) |
9 | 3,222 |
Low light (dark) |
10 | 4,986 |
| Total | 19 | 8,208 |
The scene folders follow the order in Table 14 of the paper. Scene 02 contains the low-light recording only. Each of the other nine scenes has both lighting conditions.
File layout
Match the scene, lighting condition and frame number to read an RGB–depth pair. For example, the following files come from the same normal-light frame in Scene 01.
scene01/
├── rgb/bright/000025.jpg
└── depth/bright/000025.png
Sign in to Hugging Face and request access on this page. Once approved, open the Files tab to browse the recordings.
- Downloads last month
- 75