Robust and Accurate Object Detection via Self-Knowledge Distillation

paper:https://arxiv.org/abs/2111.07239

Environments

• Python 3.7
• Cuda 10.1
• Prepare dependency

Notes: We revise MMCV to adapt adversarial algorithms. Therefore we suggest that you prepare environments strictly as follows:

conda create -n udfa python=3.7
conda activate udfa
sh prepare_env.sh

Prepare datasets

• VOC0712, download from http://host.robots.ox.ac.uk/pascal/VOC/, and place it under data directory

• COCO2017, download from https://cocodataset.org/#download, and place it under data directory

• The structure of datasets is shown as follows:

  

Train

VOC dataset

• Generate GFLV2-R34 pretrained detector (served as teacher) on PASCAL_VOC 0712:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_voc_std.py 
  cd work_dirs/gflv2_r34_fpn_voc_std
  cp epoch_12.pth ../../weights/gflv2_r34_voc_pre.pth
  

• Training GFLV2-R34 using udfa on PASCAL_VOC 0712:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_voc_kdss.py --load-from weights/gflv2_r34_voc_pre.pth
  

• Training GFLV2-R34 using udfa with advprop on PASCAL_VOC 0712:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_voc_kdss1.py --load-from weights/gflv2_r34_voc_pre.pth
  

• Training GFLV2-R34 using Det-AdvProp on PASCAL_VOC 0712:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_voc_mixbn.py --load-from weights/gflv2_r34_voc_pre.pth
  

COCO dataset

• Generate GFLV2-R34 pretrained detector (served as teacher) on COCO:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_coco_std.py 
  cd work_dirs/gflv2_r34_fpn_coco_std
  cp epoch_12.pth ../../weights/gflv2_r34_coco_pre.pth
  

• Training GFLV2-R34 using udfa on COCO:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_coco_kdss.py --load-from weights/gflv2_r34_coco_pre.pth
  

• Training GFLV2-R34 using Det-AdvProp on COCO:

  python -m torch.distributed.launch --nproc_per_node=4  train.py --launcher pytorch --config configs/gflv2/gflv2_r34_fpn_coco_mixbn.py --load-from weights/gflv2_r34_coco_pre.pth
  

Test

• Evlauate the clean AP or adversarial robustness on PASCAL_VOC 2007 test set:

  python -m torch.distributed.launch --nproc_per_node=4 test.py --launcher pytorch --configs/gflv2/gflv2_r34_fpn_voc_std.py  --checkpoint weights/gflv2_r34_voc_pre.pth --num_steps 0 --step_size 2 --eval mAP
  

• Evlauate the clean AP or adversarial robustness on COCO 2017val set:

  python -m torch.distributed.launch --nproc_per_node=4 test.py --launcher pytorch --configs/gflv2/gflv2_r34_fpn_coco_std.py  --checkpoint weights/gflv2_r34_coco_pre.pth --num_steps 0 --step_size 2 --eval bbox
  

Acknowledgement

Our project is based on ImageCorruptions, MMDetection and MMCV.