search

UMBCvision / Contextual-Adversarial-Patches

Official Repository for the CVPR 2020 AdvML Workshop paper "Role of Spatial Context in Adversarial Robustness for Object Detection"
MIT License
35 stars 5 forks source link

The AP with_fp and removed_fp is very close #2

Open xleslie opened 3 years ago

xleslie commented 3 years ago

I have try to execute bash run_pipeline_universal_patch.sh as your steps,but the AP for train with_fp is 0.8548 and the AP for train removed_fp is 0.8618.Why is the AP so close?Can you give me some suggestion?

than you~

ruixin-han commented 2 years ago

How many epochs do you need to train to run the defense program? I trained 500,000 epochs, and the ap value on the clean sample was less than 1%, and the ap value on the adversarial sample was less than 40%.