# QuadOcc

QuadOcc is a real-world dataset for 3D semantic occupancy prediction on legged robots, captured using a quadruped equipped with a Panoramic Annular Lens (PAL) camera and LiDAR. This document describes the images, point clouds, voxel labels, and supporting files in the public data package.

## 1. Dataset Overview

| Property | Description |
| --- | --- |
| Scale | 10 scenes and approximately 24,000 frames |
| Lighting | Day, dusk, and night |
| Image representations | Raw annular images and unfolded equirectangular panoramas |
| Voxel grid | `64 × 64 × 8` |
| Voxel size | `0.4 m × 0.4 m × 0.4 m` |
| Grid axes | X forward, Y left, Z up |
| Default spatial bounds | X and Y: `[-12.8, 12.8] m`; Z: `[-2.0, 1.2] m` |
| Six semantic categories | Vehicle, pedestrian, road, building, vegetation, terrain |

The spatial bounds above follow the defaults in [Appendix Section 9.7](https://arxiv.org/html/2511.03571v2#S9.SS7). Check the accompanying data configuration when loading a specific release.

## 2. Directory Structure

```text
QuadOcc/
├── sequences/
│   ├── 2024-06-20-20-46-59/
│   │   ├── images_raw/
│   │   ├── images_unfold/
│   │   ├── lidar/
│   │   ├── lidar_input/
│   │   ├── unlabel/
│   │   ├── voxel_final/
│   │   ├── PAL_intrinsic_calib_results.txt
│   │   ├── pcd_position.txt
│   │   └── timestamp_relative.txt
│   └── .../
├── maps_370.npz
├── PAL_intrinsic_calib_results.txt
└── README.md
```


## 3. Root-Level Files

| Name | Description |
| --- | --- |
| `sequences/` | Data directories for individual capture sequences. |
| `maps_370.npz` | Supporting data in NPZ format. Array keys, shapes, and purpose must be determined from the file contents and accompanying loader. |
| `PAL_intrinsic_calib_results.txt` | PAL camera intrinsics stored at the package root. |
| `README.md` | Dataset directory and file documentation. |

## 4. Per-Sequence Files

All paths below are relative to `sequences/<sequence_id>/`.

### 4.1 Images

| Directory | Contents |
| --- | --- |
| `images_raw/` | Original annular images captured by the PAL camera. |
| `images_unfold/` | Panoramic images unfolded using the PAL calibration parameters. |

The paper uses the Taylor/OCam camera model to unfold annular images into equirectangular panoramas. See [Appendix Section 9.1](https://arxiv.org/html/2511.03571v2#S9.SS1).

### 4.2 Point Clouds and Voxels

| Directory | Contents |
| --- | --- |
| `lidar/` | LiDAR scan point clouds. |
| `lidar_input/` | Binary voxel inputs constructed from LiDAR points inside the bounding volume. Stored as `.npy` arrays with shape `(64, 64, 8)`: `1` indicates that a voxel contains points, and `0` indicates that it contains no points. |
| `unlabel/` | Locations occupied in the LiDAR scan representation that have no annotation in the final semantic labels. |
| `voxel_final/` | Final semantic voxel labels with grid shape `(64, 64, 8)`. |

`lidar_input/` represents scan-point occupancy, `voxel_final/` provides semantic supervision, and `unlabel/` identifies unannotated locations. A binary input value of `0` only means that no scan point falls inside the voxel; it does not establish free space. Unannotated locations must not automatically be treated as free space either.

The six semantic category names are `vehicle`, `pedestrian`, `road`, `building`, `vegetation`, and `terrain`. Their listing order does not define numeric label IDs. Use the accompanying label mapping and loading configuration for class IDs, empty-voxel encoding, and ignore values. An example configuration is provided as `quadocc.yaml` in the release.

### 4.3 Timestamps, Poses, and Camera Intrinsics

| File | Contents |
| --- | --- |
| `timestamp_relative.txt` | Relative timestamps for the sequence. |
| `pcd_position.txt` | Point-cloud poses synchronized with the sequence data. |
| `PAL_intrinsic_calib_results.txt` | PAL camera intrinsics stored within the sequence directory. |

Check image, point-cloud, and label correspondence using timestamps and the loading configuration. Pose column order, rotation representation, transform direction, and selection between root-level and sequence-level calibration files depend on the actual files and accompanying code.

## 5. Semantic Label Construction

The labels are constructed using panoramic semantic segmentation, image-to-point semantic transfer, multi-frame point-cloud aggregation, and manual corrections.

## 6. References

- [OneOcc: Semantic Occupancy Prediction for Legged Robots with a Single Panoramic Camera](https://arxiv.org/abs/2511.03571)
- [Appendix: QuadOcc Dataset Construction](https://arxiv.org/html/2511.03571v2#S9)
- [Official OneOcc Repository](https://github.com/MasterHow/OneOcc)

## 7. Citation

If you use the QuadOcc dataset in your research, please cite OneOcc. Thank you!

```bibtex
@InProceedings{Shi_2026_CVPR,
  author    = {Shi, Hao and Wang, Ze and Guo, Shangwei and Duan, Mengfei and Wang, Song and Chen, Teng and Yang, Kailun and Wang, Lin and Wang, Kaiwei},
  title     = {{OneOcc}: Semantic Occupancy Prediction for Legged Robots with a Single Panoramic Camera},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2026},
  pages     = {14229--14240}
}
```