Effective Feature Enhancement and Model Ensemble Strategies in Tiny Object Detection
Abstract
We introduce a novel tiny-object detection network that achieves better accuracy than existing detectors on TinyPerson dataset. It is an end-to-end detection framework developed on PaddlePaddle. A suit of strategies are developed to improve the detectors performance including: 1) data augmentation based on scale-match that aligns the object scales between the existing large-scale dataset and TinyPerson; 2) comprehensive training methods to further improve detection performance by a large margin; 3) model refinement based on the enhanced PAFPN module to fully utilize semantic information; 4) a hierarchical coarse-to-fine ensemble strategy to improve detection performance based on a well-designed model pond.
Cite
Text
Feng et al. "Effective Feature Enhancement and Model Ensemble Strategies in Tiny Object Detection." European Conference on Computer Vision Workshops, 2020. doi:10.1007/978-3-030-68238-5_24Markdown
[Feng et al. "Effective Feature Enhancement and Model Ensemble Strategies in Tiny Object Detection." European Conference on Computer Vision Workshops, 2020.](https://mlanthology.org/eccvw/2020/feng2020eccvw-effective/) doi:10.1007/978-3-030-68238-5_24BibTeX
@inproceedings{feng2020eccvw-effective,
title = {{Effective Feature Enhancement and Model Ensemble Strategies in Tiny Object Detection}},
author = {Feng, Yuan and Wang, Xiaodi and Xin, Ying and Zhang, Bin and Liu, Jingwei and Mao, Mingyuan and Xu, Sheng and Zhang, Baochang and Han, Shumin},
booktitle = {European Conference on Computer Vision Workshops},
year = {2020},
pages = {324-330},
doi = {10.1007/978-3-030-68238-5_24},
url = {https://mlanthology.org/eccvw/2020/feng2020eccvw-effective/}
}