This repository hosts the Human-Parts dataset used in our DID-Net (Detector-in-Detector Network). Our work, presented at ACCV 2018, introduces a novel framework that enhances detection performance by leveraging the inherent correlation between the human body and its parts. We utilize a region-based detection structure with dual detectors that focus on the human body and body parts in a hierarchical manner.
The Human-Parts dataset, essential for training our model, comprises 14,962 images and 106,879 annotations, targeting various aspects of the human form, such as the body, face, and hands. We invite researchers to utilize this dataset for related tasks and applications in computer vision.
- Multi-level object detection framework
- Coarse-to-fine analysis via dual detectors
- Trained end-to-end on a newly collected and labeled dataset
- Demonstrates significant improvements in detecting multi-level human-related objects
If you find the Human-Parts dataset or DID-Net useful in your research, please consider citing our paper:
@inproceedings{li2019detector,
title={Detector-in-detector: Multi-level analysis for human-parts},
author={Li, Xiaojie and Yang, Lu and Song, Qing and Zhou, Fuqiang},
booktitle={Proceedings of the Asian Conference on Computer Vision},
pages={228--240},
year={2019},
organization={Springer}
}For more details, refer to the full paper on arXiv.
Access the Human-Parts dataset through the following links:
This project is now maintained within this repository. For further inquiries or issues, please contact [email protected] or leave an issue in this repository.