---
title: referring_vos_toydata
canonical_url: "https://www.modelscope.cn/datasets/modelscope/referring_vos_toydata"
md_url: "https://www.modelscope.cn/datasets/modelscope/referring_vos_toydata.md"
repository: modelscope/referring_vos_toydata
chinese_name: "文本指导的视频目标分割数据集-toydata"
last_updated: 2022-11-09
license: "Apache License 2.0"
storage_size: "3.1 KB"
domain:
  - video
tasks:
  - Movie-Scene-Segmentation
encoding:
  - mp4
downloads: 382
stars: 0
---

# referring_vos_toydata

> referring_vos_toydata - modelscope 在 ModelScope 开源的数据集。用于文本指导的视频目标分割数据集

modelscope/referring_vos_toydata 是 ModelScope 魔搭社区上的Movie-Scene-Segmentation数据集，涉及 video 领域，存储大小 3.1 KB，采用 Apache License 2.0 许可。

- **Repository**: modelscope/referring_vos_toydata
- **License**: Apache License 2.0
- **Tasks**: Movie-Scene-Segmentation
- **Domain**: video
- **Storage size**: 3.1 KB
- **Downloads**: 382
- **Stars**: 0
- **Last updated**: 2022-11-09

Source: https://www.modelscope.cn/datasets/modelscope/referring_vos_toydata

---

## 数据集整体描述
用于文本指导的视频目标分割任务的数据集。
本数据集视频文件来自于[A2D数据集](https://web.eecs.umich.edu/~jjcorso/r/a2d/)，
标注文件来自于[kgavrilyuk](https://kgavrilyuk.github.io/publication/actor_action/)。

**本仓库只提供了其中5个视频相关的文件进行简单测试，可以从上述两个链接中获取完整的数据集。**

### 数据集简介
该数据集包含3782个视频，其中3036个视频用于训练，746个视频用于测试。
数据集中视频的帧数在24帧到332帧不等，其平均帧数为136帧。

数据集包含7个actor类别：adult、baby、ball、bird、car、cat和dog以及8个action类别：climb、crawl、eat、fly、jump、roll、run和walk。

[kgavrilyuk](https://kgavrilyuk.github.io/publication/actor_action/)提供的标注文件中，新增了针对每个actor-action元组的描述文本和对应的intance id。


## 数据集的格式和结构
### 数据格式
本仓库不包含完整的原始数据集文件，仅仅提供了5个样例视频。其数据格式与原始数据集格式保持一致，具体格式如下：
```
referring_vos_toydata/ 
    ├── Release/ (provided by origin A2D dataset)
    │   ├── videoset.csv  (videos metadata file)
    │   └── CLIPS320/
    │       └── *.mp4     (video files)
    └── text_annotations/ (provided by kgavrilyuk)
        ├── a2d_annotation.txt  (actual text annotations)
        ├── a2d_missed_videos.txt
        └── a2d_annotation_with_instances/ 
            └── */ (video folders)
                └── *.h5 (annotations files) 
```

### 数据集加载方式
```Python
from modelscope.msdatasets import MsDataset
ms_ds_train = MsDataset.load('referring_vos_toydata', split='train')
print(ms_ds_train._hf_ds.config_kwargs)
assert next(iter(ms_ds_train.config_kwargs['split_config'].values()))
```



## 引用方式
```
@inproceedings{XuHsXiCVPR2015,
  author = {Xu, C. and Hsieh, S.-H. and Xiong, C. and {\bf Corso}, {\bf J. J.}},
  booktitle = {{Proceedings of IEEE Conference on Computer Vision and Pattern Recognition}},
  datadownload = {http://web.eecs.umich.edu/~jjcorso/r/a2d},
  poster = {http://web.eecs.umich.edu/~jjcorso/pubs/xu_corso_CVPR2015_A2D_poster.pdf},
  tags = {computer vision, activity recognition, video understanding, semantic segmentation},
  title = {Can Humans Fly? {Action} Understanding with Multiple Classes of Actors},
  url = {http://web.eecs.umich.edu/~jjcorso/pubs/xu_corso_CVPR2015_A2D.pdf},
  year = {2015}
}

@inproceedings{gavrilyuk2018actor,
  title={Actor and action video segmentation from a sentence},
  author={Gavrilyuk, Kirill and Ghodrati, Amir and Li, Zhenyang and Snoek, Cees GM},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={5958--5966},
  year={2018}
}
```