Skip to content

Latest commit

 

History

History
 
 

ddod

README.md

Disentangle Your Dense Object Detector

Introduction

This directory contains the configs and results of Disentangle Your Dense Object Detector.

Results and Models

Method Backbone Lr schd box AP box AP by authors Config Download
ATSS (IoU) R-50 1x - 39.4 config -
ATSS+DDOD R-50 1x 41.3 41.6 config model
  • DDOD adopts a variant of ATSS as a baseline, which predicts IoU instead of centerness.
  • We use a total batch size of 16 (in 4 GPUs) with lr=0.01, while the authors use a total batch size of 32 (in 8 GPUs) with lr=0.02.
  • Other possible reasons for AP difference are randomness, instability due to fp16, and implementation change of mixed-precision training (MMCV >= 1.3.2).

Citation

@article{chen2021disentangle,
  title={Disentangle Your Dense Object Detector},
  author={Zehui Chen and Chenhongyi Yang and Qiaofei Li and Feng Zhao and Zhengjun Zha and Feng Wu},
  journal={arXiv preprint arXiv:2107.02963},
  year={2021}
}