XS-VID: An Extremely Small Video Object Detection Dataset

Paper: XS-VID: A Large-Scale Benchmark for Small Object Detection and Tracking in Videos
Code: YOLOFT

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 huggingface using command-line tools in both Linux and Windows environments.

Prerequisites

• Python and pip: Ensure Python and pip are installed on your system.

• huggingface Library: Install the huggingface_hub library using pip:

  pip install huggingface_hub
  

Download and Extract Dataset

Linux Command:

pip install huggingface_hub && \
huggingface-cli download lanlanlan23/XS-VID --repo-type dataset --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):

pip install huggingface_hub && ^
huggingface-cli download lanlanlan23/XS-VID --repo-type dataset --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:

   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 [email protected].