---
title: Sekai2_Panoramics
canonical_url: "https://www.modelscope.ai/datasets/kangverse/Sekai2_Panoramics"
md_url: "https://www.modelscope.ai/datasets/kangverse/Sekai2_Panoramics.md"
repository: kangverse/Sekai2_Panoramics
last_updated: 2026-09-09
license: other
storage_size: "1.4 TB"
downloads: 56905
stars: 8
---

# Sekai2_Panoramics

> Sekai2_Panoramics - An open-source dataset by kangverse on ModelScope. Sekai2 Panoramics — 20-hour public subset

kangverse/Sekai2_Panoramics is a dataset on ModelScope. totalling 1.4 TB. licensed under other.

- **Repository**: kangverse/Sekai2_Panoramics
- **License**: other
- **Storage size**: 1.4 TB
- **Downloads**: 56905
- **Stars**: 8
- **Last updated**: 2026-09-09

Source: https://www.modelscope.ai/datasets/kangverse/Sekai2_Panoramics

---

# Sekai2 Panoramics — 20-hour public subset

This repository contains the full-resolution videos and paired camera-pose,
caption, and camera-model annotations for the 20-hour panoramic subset of
**Sekai2**. See the
paper: **[Sekai2: From World Exploration to Interactive World Modeling](https://arxiv.org/abs/2608.09449)**.

## Contents

The release contains 173 complete 360° equirectangular sequences totaling
20.105 hours:

```text
videos/<english_stem>.mp4
annotations/pose/<english_stem>.npz
annotations/caption/<english_stem>.json
annotations/intrinsic/<english_stem>.npz
annotations/intrinsic/<english_stem>_camera.txt
metadata/panoramic_public_20h.csv
metadata/annotation_checksums.csv
```

`metadata/panoramic_public_20h.csv` is the canonical index. Its `pose_file`,
`caption_file`, `intrinsic_file`, and `camera_type_file` columns are
repository-relative paths and use the same English stem as
`english_filename`. All 173 CSV rows have one video and a complete matching
annotation set. `metadata/annotation_checksums.csv` provides SHA-256 checksums
for all four released annotation files.

The sequences use a spherical `PANORAMA` camera model rather than a pinhole
camera matrix. Accordingly, each intrinsic NPZ stores frame indices in `inds`
and a zero-valued `(N, 4)` sentinel array in `data`; the matching
`_camera.txt` identifies every frame as `PANORAMA`.

These panoramic videos are released whole and must not be timestamp-cut.

## Where the three Sekai2 components live

| Component | Location |
|---|---|
| `sekai` perspective data | [Hugging Face: Kangverse/Sekai2_Real_World](https://huggingface.co/datasets/Kangverse/Sekai2_Real_World) |
| `sekai2` perspective data, including the static supplement | [Hugging Face: Kangverse/Sekai2_Real_World](https://huggingface.co/datasets/Kangverse/Sekai2_Real_World) |
| 20-hour panoramic subset | This ModelScope repository |

## Citation

```bibtex
@article{he2026sekai2,
  title   = {Sekai2: From World Exploration to Interactive World Modeling},
  author  = {He, Kang and Peng, Wenshuo and Gao, Zihui and Tan, Jiaming and
             Zhang, Kaipeng and Ge, Yongtao},
  journal = {arXiv preprint arXiv:2608.09449},
  year    = {2026}
}
```
