Generate Adversarial Attacked Image for Object Detection models using MMDetection

• You should put the files on /mmdetection/{here}
  • This code use mmdetection 3.x version. DO NOT USE 2.x
• In this version, you can use only DAG method. I might update another methos later.
• This method very sensitive to changes in pixel values, so it stores the attacked image in [0,1] format rather than [0,255] format. For this purpose, the attacked images are stored in the form of a numpy array(.npy).

• This pipeline follows the following scheme
  1. Import a file with a video cropped to an image.
  2. Create a Pseudo Ground Truth(GT) of the file using Model Predcition.
  3. Create a Fake GT by randomly changing the label value of the model's Pseudo GT.
  4. Create an adversarial image using the fake GT.
  5. Save the image with Detection Visualization.