README.md

---
pretty_name: PanoCARLA
annotations_creators:
- machine-generated
task_categories:
- depth-estimation
- image-segmentation
tags:
- panorama-understanding
- panoramic
- video
- carla
- depth
- segmentation
size_categories:
- 100K<n<1M
license: cc-by-4.0
---

# PanoCARLA

PanoCARLA is a synthetic panoramic video dataset generated with CARLA 0.9.16 for panoramic understanding tasks. It provides temporally aligned equirectangular RGB images, metric depth annotations, semantic segmentation, and camera poses across multiple CARLA towns and roaming routes.

PanoCARLA has been used for the training of [PVDepth](https://github.com/ChuanxinSong/PVDepth).

## Dataset Overview

PanoCARLA contains 126,696 frames collectedin eight CARLA towns. The PVDepth protocol uses six towns for training and two disjoint towns for evaluation.

| Split | Towns | Routes | Source clips | Frames |
|---|---|---:|---:|---:|
| Train | Town01, Town03, Town04, Town05, Town06, Town07 | 54 | 95 | 112,988 |
| Test | Town02, Town10 | 13 | 22 | 13,708 |
| **Total** | **8 towns** | **67** | **117** | **126,696** |

### Per-town Statistics

| Town | Routes | Clips | Frames |
|---|---:|---:|---:|
| Town01 | 8 | 18 | 15,512 |
| Town02 | 6 | 14 | 6,626 |
| Town03 | 8 | 22 | 17,822 |
| Town04 | 8 | 13 | 23,531 |
| Town05 | 10 | 19 | 25,436 |
| Town06 | 8 | 9 | 15,788 |
| Town07 | 12 | 14 | 14,899 |
| Town10 | 7 | 8 | 7,082 |

## Repository Structure

```text
PanoCARLA/
├── panocarla_h5/                  # Full train/test dataset in HDF5 format (2400 x 1200)
├── town0210_res1024_512.tar       # Raw Town02/Town10 benchmark data (1024 x 512)
├── test_benchmark/                # Benchmark JSON files at different sampling rates
├── setting_dynamic.json           # Full clip and frame index
├── statistics_report_dynamic.txt  # Per-town statistics
├── carla_path_vis/                # Roaming-route visualizations
└── vis_mp4/                       # Panoramic preview videos for all captured routes
```

The complete dataset is distributed through `panocarla_h5/`. The raw PNG/NPY representation is provided only for the Town02/Town10 evaluation split in `town0210_res1024_512.tar`.

`setting_dynamic.json` indexes all towns and retains paths from the original raw-data layout. For training towns, use the HDF5 files rather than expecting the referenced raw PNG/NPY files to be present.

## HDF5 Format

Each file in `panocarla_h5/` stores one continuous source clip. The first dimension `T` is the number of frames in that clip.

| Key | Shape | dtype | Description |
|---|---|---|---|
| `rgb` | `[T, 1200, 2400, 3]` | `uint8` | Equirectangular RGB frames |
| `depth` | `[T, 1200, 2400, 1]` | `float16` | Metric panoramic depth in meters |
| `seg` | `[T, 1200, 2400, 1]` | `uint8` | CARLA semantic segmentation IDs |
| `poses` | `[T, 4]` | `float32` | Camera poses in `(x, y, z, yaw)` order |

The position components `x`, `y`, and `z` are expressed in meters in the CARLA world coordinate system, and `yaw` is expressed in degrees. Pitch and roll are always zero and are therefore not stored.

Semantic IDs follow the official [CARLA 0.9.16 semantic segmentation specification](https://carla.readthedocs.io/en/0.9.16/ref_sensors/#semantic-segmentation-camera).

### Reading an HDF5 Clip

```python
import h5py

h5_path = "panocarla_h5/town01_path1_town01_path1_clip_0.h5"

with h5py.File(h5_path, "r") as file:
    rgb = file["rgb"][0]           # [1200, 2400, 3], uint8
    depth = file["depth"][0]       # [1200, 2400, 1], float16, meters
    segmentation = file["seg"][0] # [1200, 2400, 1], uint8
    frame_id = file["frame_ids"][0]
    pose = file["poses"][0]        # x, y, z, yaw
```

For large clips, read only the required frame indices instead of loading the entire file into memory.

## Benchmark Split

Town02 and Town10 form the evaluation split. The raw benchmark data is stored in `town0210_res1024_512.tar`, and `test_benchmark/` provides JSON configurations at 2, 10, and 20 FPS with sequence lengths of 50, 90, and 110 frames, respectively.

## Download

This is a large gated dataset. Request access on the dataset page and authenticate with the Hugging Face CLI before downloading:

```bash
hf auth login
```

Download the full HDF5 dataset:

```bash
hf download Soon122/PanoCARLA \
  --repo-type dataset \
  --include "panocarla_h5/*" \
  --include "setting_dynamic.json" \
  --local-dir PanoCARLA
```

Download only the evaluation data:

```bash
hf download Soon122/PanoCARLA \
  town0210_res1024_512.tar \
  --repo-type dataset \
  --local-dir PanoCARLA

hf download Soon122/PanoCARLA \
  --repo-type dataset \
  --include "test_benchmark/*" \
  --local-dir PanoCARLA

tar -xf PanoCARLA/town0210_res1024_512.tar -C PanoCARLA
```

## License

PanoCARLA is licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), which permits commercial use with attribution.

## Citation

If you use PanoCARLA in your research, please cite the PVDepth paper:

```bibtex
@inproceedings{song2026pvdepth,
  author    = {Song, Chuanxin and Peng, Peixi},
  title     = {PVDepth: Panoramic Video Depth Estimation via Geometry-Aware Spatiotemporal Adaptation},
  booktitle = {ICML},
  year      = {2026}
}
```

## Acknowledgements

PanoCARLA was generated with [CARLA 0.9.16](https://github.com/carla-simulator/carla). We thank the CARLA team for making the simulator publicly available.