hciilab / scut-head-dataset-release

SCUT HEAD is a large-scale head detection dataset, including 4405 images labeld with 111251 heads.

News

2018-1-16 SCUT-HEAD v1.0 is released with images and head bounding box annotations.

Description

SCUT-HEAD is a large-scale head detection dataset, including 4405 images labeld with 111251 heads. The dataset consists of two parts. PartA includes 2000 images sampled from monitor videos of classrooms in an university with 67321 heads annotated. PartB includes 2405 images crawled from Internet with 43930 heads annotated. We have labelled every visable heads with xmin, ymin, xmax and ymax coordinates and ensured that annotations cover the entire head including the blocked parts but without extra background. Both PartA and PartB are divided into training and testing parts. Our dataset follows the standard of Pascal VOC.

PartA

PartA includes 2000 images sampled from monitor videos of classrooms in an university with 67321 heads annotated. 1500 images of PartA are for training and 500 for testing. Because classrooms in an university usually looks similar and the poses of people vary less, we carefully choose representative images to gain variance and reduce similarity. Representative image and annotations and the histogram of people counts are shown below.

PartB

PartB includes 2405 images with 43940 heads annotated. 1905 images of PartB are for training and 500 for testing. The images are crawled from Internet. The urls of images are also provided in the dataset. Representative image and annotations and the histogram of people counts are shown below.

Download

PartA of SCUT-HEAD [Google Drive] [Baidu Drive]

PartB of SCUT-HEAD [Google Drive] [Baidu Drive]

Related Datasets

HollywoodHead dataset HolleywoodHeads dataset is a head detection datset. HollywoodHeads dataset contains 369846 human heads annotated in 224740 video frames from 21 Hollywood movies.

Brainwash dataset Brainwash dataset is related for face detection. Brainwash dataset contains 11917 images with 91146 labeled people.

Citation

@article{peng2018detecting,
  title={Detecting Heads using Feature Refine Net and Cascaded Multi-scale Architecture},
  author={Peng, Dezhi and Sun, Zikai and Chen, Zirong and Cai, Zirui and Xie, Lele and Jin, Lianwen},
  journal={arXiv preprint arXiv:1803.09256},
  year={2018}
}

Remark

The SCUT-HEAD database is free to the academic community for research purpose usage only.

Contact

For any questions about this database please contact the authors by sending email to [email protected] and [email protected].

scut-head-dataset-release's Issues

Detecting Heads using Feature Refine Net and Cascaded Multi-scale Architecture

Hello, could you release your code, I really want to implement your work.

关于代码是否开源的问题

是否考虑开源相关部分的代码呢，这将对我们的研究有所帮助，十分感谢

How many epoches have you trained?

Citation paper

Hi~, I am wondering where could I find your corresponding paper ?

Incorrect VOC format

There are some issues with the XML files, e.g. the image filename/path being incorrect, and I seem to recall a colleague noticing that the image sizes in the XML didn't always match the actual image sizes (and were sometimes 0 maybe?). I've also noticed some of the bounding boxes are invalid (i.e. have xmin==xmax, and ymin==ymax).

Would it be possible to fix these?

Have you ever released your code? If you would do that, please let me know.

some incomplete or occluded heads without annotation

After training my person head detection model using this dataset, I found that many incomplete or occluded person heads were not detected. Upon further investigation, I discovered that some cases in the dataset are also lacking annotations. How much impact would this missing data have on my model?
without annotation examples: