---
license: other
task_categories:
- image-to-image
tags:
- novel-view-synthesis
- autonomous-driving
- lidar
- multi-view
size_categories:
- 10K<n<100K
---

# MHDCSM/cv_ds_all

WebDataset-format shards of the autonomous driving novel-view-synthesis dataset.

Splits:
- `train-*.tar` — training samples (includes ground-truth target images)
- `test-*.tar`  — test samples (no targets)

Each tar holds many samples. Per-sample files share a key (the `sample_id`):

| Suffix | Content |
|--------|---------|
| `meta.json` | camera poses, intrinsics, timestamps, target camera name |
| `t0.<cam>.jpg`, `t1.<cam>.jpg` | 6 camera views per timestamp |
| `lidar.npz` | dense world-frame point cloud (xyz, intensity) |
| `target.<cam>.jpg` | ground-truth target image (train only) |

## Streaming usage

```python
import webdataset as wds
url = "https://huggingface.co/datasets/MHDCSM/cv_ds_all/resolve/main/train-{000000..NNNNNN}.tar"
ds = wds.WebDataset(url).decode()
for sample in ds:
    meta = sample["meta.json"]
    img  = sample["t0.front.jpg"]
```