HRDNet: High-Resolution Detection Network for Small Objects

Ziming Liu1,3 · Guangyu Gao1,✉ · Lin Sun2 · Zhiyuan Fang1

1Beijing Institute of Technology · 2Magic Leap, Inc. · 3Inria, Université Côte d'Azur

ICME 2021 · oral presentation

HRDNet architecture with multi-depth image pyramid and multi-scale feature pyramid
High-resolution inputs are fed to shallow backbones to preserve positional detail, while low-resolution inputs go through deep backbones for semantics. MS-FPN then aligns and fuses the resulting multi-scale feature groups.

Abstract

Small object detection is a very challenging yet practical vision task. With deep network-based methods, the contextual information of small objects may disappear when the network goes deeper. An intuitive solution is to increase the input resolution; however, this aggravates the large variance of object scale and introduces unbearable computation cost.

To leverage the benefits of high-resolution images without bringing up new problems, we propose a High-Resolution Detection Network (HRDNet) which takes multiple resolution inputs with multi-depth backbones. We propose the Multi-Depth Image Pyramid Network (MD-IPN) and the Multi-Scale Feature Pyramid Network (MS-FPN). MD-IPN maintains multiple position information using multiple depth backbones: high-resolution input is fed into a shallow network to reserve more positional information and reduce computational cost, while low-resolution input is fed into a deep network to extract more semantics. MS-FPN aligns and fuses the multi-scale feature groups generated by MD-IPN to reduce information imbalance.

Extensive experiments are conducted on COCO2017 and on the typical small object dataset VisDrone 2019. HRDNet achieves state-of-the-art results on both datasets with significant improvements on small objects.

Paper

First page of the HRDNet paper

Ziming Liu, Guangyu Gao, Lin Sun, Zhiyuan Fang.
HRDNet: High-Resolution Detection Network for Small Objects.
IEEE International Conference on Multimedia and Expo (ICME), 2021.

BibTeX 
@inproceedings{liu2021hrdnet,
  title        = {HRDNet: high-resolution detection network for small objects},
  author       = {Liu, Ziming and Gao, Guangyu and Sun, Lin and Fang, Zhiyuan},
  booktitle    = {IEEE International Conference on Multimedia and Expo (ICME)},
  pages        = {1--6},
  year         = {2021},
  organization = {IEEE}
}

Acknowledgements

Supported mainly by the National Natural Science Foundation of China under Grant 61972036, and in part by grants under Grant 91746210.