XS-VID

# XS-VID: An Extremely Small Video Object Detection Dataset

## Dataset Description

XS-VID is designed as a benchmark dataset for extremely small video object detection. It is intended to evaluate the performance of video object detection models, particularly focusing on efficiency and effectiveness in resource-limited situations. The dataset includes a variety of videos and scenarios to comprehensively assess model capabilities.

**[News]**: XS-VIDv2 is coming soon! We are excited to announce the upcoming release of XS-VIDv2, which will feature a significantly expanded dataset with many new videos and scenarios. Stay tuned for updates!

To access the XS-VID benchmark go to **https://gjhhust.github.io/XS-VID/**

## Dataset Download

### Using Command Line

This guide provides instructions for downloading and extracting the XS-VID dataset from ModelScope using command-line tools in both Linux and Windows environments.

#### Prerequisites

* **Python and pip:** Ensure Python and pip are installed on your system.
* **ModelScope Library:** Install the ModelScope library using pip:

```bash
pip install modelscope
```

#### Download and Extract Dataset

**Linux Command:**

```bash
pip install modelscope && \
modelscope download --dataset lanlanlanrr/XS-VID --local_dir ./XS-VID && \
mkdir -p ./XS-VID/{annotations,images} && \
unzip -o ./XS-VID/annotations.zip -d ./XS-VID/annotations && \
find ./XS-VID -name 'videos_subset_*.zip' -exec unzip -o {} -d ./XS-VID/images \; && \
rm -f ./XS-VID/*.zip
```

**Windows Command (CMD):**

```bash
pip install modelscope && ^
modelscope download --dataset lanlanlanrr/XS-VID --local_dir ./XS-VID && ^
mkdir "./XS-VID\annotations" && mkdir "./XS-VID\images" && ^
powershell -Command "Expand-Archive -Path './XS-VID/annotations.zip' -DestinationPath './XS-VID/annotations' -Force" && ^
for /r "./XS-VID" %f in (videos_subset_*.zip) do powershell -Command "Expand-Archive -Path '%f' -DestinationPath './XS-VID/images' -Force" && ^
del /f /q "./XS-VID\*.zip"
```

### Expected Folder Structure

After running the download and extraction commands, the XS-VID dataset folder should have the following structure:

```
./XS-VID/
├── annotations/ # Annotation files
└── images/ # Video frames (extracted from videos_subset_*.zip)
```

### Notes

* The download script automatically deletes ZIP files after successful extraction.
* Ensure you have sufficient disk space available (approximately the size of the ZIP files plus the extracted content).

## Evaluation Tool Usage

To evaluate your models on the XS-VID dataset, please follow these steps:

1. **Clone the repository:** Obtain the evaluation tool files, including `eval_tool.py`, `cocoeval.py`, and `mask.py` from the main branch of the XS-VID repository.
2. **Set JSON paths:** In `eval_tool.py`, configure the paths to your test COCO JSON annotation file and prediction JSON file.
3. **Run evaluation:** Execute the evaluation script using the command:

```bash
python eval_tool.py
```

## Citation

If you utilize the XS-VID dataset in your research or applications, please cite the following paper:

```
@article{guo2024XSVID,
title={XS-VID: An Extremely Small Video Object Detection Dataset},
author={Jiahao Guo, Ziyang Xu, Lianjun Wu, Fei Gao, Wenyu Liu, Xinggang Wang},
journal={arXiv preprint arXiv:2407.18137},
year={2024}
}
```

## Support and Contact

For any questions or issues regarding the XS-VID benchmark, please feel free to contact us at [e-mail].
```

# XS-VID：极小型视频目标检测数据集

## 数据集说明
XS-VID 专为极小型视频目标检测（video object detection）任务设计，作为基准数据集（benchmark dataset）用于评测视频目标检测模型的性能，尤其聚焦于资源受限场景（resource-limited situations）下的模型效率与效果。该数据集涵盖多元视频与应用场景，可全面评估模型的综合能力。

**【动态】**：XS-VIDv2 即将发布！我们很高兴地宣布，XS-VIDv2 即将正式推出，其数据集规模将大幅扩充，新增大量视频与场景。敬请关注后续更新！

如需访问XS-VID基准数据集，请前往 **https://gjhhust.github.io/XS-VID/**

## 数据集下载
### 命令行下载方式
本指南提供在Linux与Windows环境下，通过命令行工具从ModelScope下载并解压XS-VID数据集的操作步骤。

#### 前置要求
* **Python与pip**：确保系统中已安装Python与pip工具。
* **ModelScope库**：通过pip安装ModelScope库：
bash
pip install modelscope

#### 下载与解压数据集
**Linux命令：**
bash
pip install modelscope &&
modelscope download --dataset lanlanlanrr/XS-VID --local_dir ./XS-VID &&
mkdir -p ./XS-VID/{annotations,images} &&
unzip -o ./XS-VID/annotations.zip -d ./XS-VID/annotations &&
find ./XS-VID -name 'videos_subset_*.zip' -exec unzip -o {} -d ./XS-VID/images ; &&
rm -f ./XS-VID/*.zip

**Windows命令（CMD）：**
bash
pip install modelscope && ^
modelscope download --dataset lanlanlanrr/XS-VID --local_dir ./XS-VID && ^
mkdir "./XS-VIDannotations" && mkdir "./XS-VIDimages" && ^
powershell -Command "Expand-Archive -Path './XS-VID/annotations.zip' -DestinationPath './XS-VID/annotations' -Force" && ^
for /r "./XS-VID" %f in (videos_subset_*.zip) do powershell -Command "Expand-Archive -Path '%f' -DestinationPath './XS-VID/images' -Force" && ^
del /f /q "./XS-VID*.zip"

### 预期文件夹结构
执行下载与解压命令后，XS-VID数据集文件夹应具备如下结构：

./XS-VID/
├── annotations/ # 标注文件
└── images/ # 视频帧（从videos_subset_*.zip中提取）

### 注意事项
* 下载脚本将在解压完成后自动删除压缩包文件。
* 请确保磁盘空间充足，所需空间约为压缩包体积与解压后内容体积之和。

## 评测工具使用说明
若需在XS-VID数据集上评测模型性能，请遵循以下步骤：
1. **克隆仓库**：从XS-VID仓库的主分支获取评测工具文件，包括`eval_tool.py`、`cocoeval.py`与`mask.py`。
2. **配置JSON路径**：在`eval_tool.py`中，配置测试集COCO格式标注JSON文件与预测结果JSON文件的路径。
3. **运行评测**：通过如下命令执行评测脚本：
bash
python eval_tool.py

## 引用格式
若您在研究或应用中使用XS-VID数据集，请引用如下论文：

@article{guo2024XSVID,
title={XS-VID: An Extremely Small Video Object Detection Dataset},
author={Jiahao Guo, Ziyang Xu, Lianjun Wu, Fei Gao, Wenyu Liu, Xinggang Wang},
journal={arXiv preprint arXiv:2407.18137},
year={2024}
}

## 支持与联系
若您对XS-VID基准数据集有任何疑问或问题，请联系我们：[e-mail].

搜集汇总

背景与挑战

背景概述

XS-VID是一个专为极小型视频目标检测设计的基准数据集，用于评估模型在资源有限情况下的性能。该数据集包含多种视频和场景，总大小为29.60GB，采用CC-BY-4.0许可证，并计划推出扩展版本XS-VIDv2。