- This is the official implementation for "Relevance Attack on Detectors" published in Pattern Recognition.
conda create -n rad python=3.6.2
conda activate rad
conda install tensorflow-gpu==1.13.1
pip install pillow==6.2.1 opencv-python==4.1.2.30 scikit-image==0.15.0 matplotlib==2.2.2
- COCO/val2017 download
- COCO/annotations/instances_val2017_resized.json
python yolov3.py 0
- Run customized parameters on GPU 1
python yolov3.py 1 --alpha=1 --epsilon=10 --num_iteration=5
- model_data/resnet50_coco_best_v2.1.0.h5 download
- install keras-retinanet
- Run default parameters on GPU 0
python retinanet.py 0
- Run customized parameters on GPU 1
python retinanet.py 1 --alpha=1 --epsilon=10 --num_iteration=5
- pytorch 1.3.1
- install mmdetection 1.1.0
- models download
- Test on GPU0 for 2020-06-09-15-19-07_RAD_SI_YOLOv3_Index0to5000_Eps2_Iter10
python test.py 2020-06-09-15-19-07_RAD_SI_YOLOv3_Index0to5000_Eps2_Iter10 0
- Adversarial Objects in COntext (AOCO) is the first adversarial dataset for object detection and instance segmentation.
- AOCO serves as a potential benchmark to evaluate the robustness of detectors, which is beneficial to network designers.
- AOCO is generated from the full COCO 2017 validation set with 5k samples.
- AOCO contains 5K adversarial samples for evaluating object detection and 5K for instance segmentation.
- All 10K samples in AOCO are crafted by our RAD.
- [download] password: gbug