/Rscup2019-Object-Detection-Track-3th-Solution

Rscup2019 Object Detection Track 3th Solution

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

RSCUP2019 object detection track 3th solution

Introduction

2019遥感图像稀疏表征与智能分析竞赛第三名方案

Major tricks

  • Mask RCNN 0.3541
  • Hybrid Task RCNN + deform conv 0.36633
  • expand bbox 0.364
  • cascade score thresh adopt to 0.5 0.366
  • small number class augmentation 0.372
  • cross_entropy weighted 0.369
  • sync BN 0.376
  • IOU sampler 0.383
  • pesudo label fine tune 0.362
  • balanced sampler 0.369
  • augmentation 0.40
  • 3 scale test 0.399
  • resnext101 0.41
  • scale2 finetune 0.43

Installation

Please refer to INSTALL.md for installation and dataset preparation.

PrepareData

python ./tools/prepare_data.py OR Dataprepare.ipynb

data will generate in ./data/rscup/annotation/ and ./data/rscup/train

Train

./tools/dist_train.sh ./configs/rscup/htc_next_3s.py <gpu_num>

Test

./tools/dist_test.sh ./configs/rscup/htc_next_3s.py ./work_dirs/htc_next_3s/epoch*.pth <gpu_num> \

--out test.pkl

Merge_result

python ./tools/merge_result.py

Convert2caffe

We achieved the converter from Hybrid Task cascade RCNN trained with mmdetection to Caffe.

Please refer " ".