DHRNet aims to generate different scale-sensitive weights (a parallel multi-branch architecture with a soft conditional gate module) during feature fusion in HRNet, thus consistently improving the performance across different scales on Citypersons.
The code base of our work is MMDetection(Here is MMDetection-1.2.0).
Config: configs/hrnet/faster_rcnn_hrnetv2p_w18_dynamic.py
Backbone: DHRNet
@article{DHRNet,
title = {Learning a Dynamic High-Resolution Network for Multi-Scale Pedestrian Detection.},
author = {Mengyuan Ding, Shanshan Zhang, Jian Yang.},
journal = {International Conference on Pattern Recognition(ICPR)},
year={2020}
}