FCA/README.md

## FCA: Learning a 3D Full-coverage Vehicle Camouflage for Multi-view Physical Adversarial Attack

Case study of the FCA. The code can be find in [FCA](https://github.com/winterwindwang/Full-coverage-camouflage-adversarial-attack.git).

### Cases of digital attack

<table>
  <tr>
    <td> <img src = 'https://github.com/winterwindwang/Full-coverage-camouflage-adversarial-attack/blob/gh-pages/assets/distance_10_elevation_30_ori_pred.gif?raw=true'/></td>
    <td><img src = 'https://github.com/winterwindwang/Full-coverage-camouflage-adversarial-attack/blob/gh-pages/assets/distance_10_elevation_30_adv_pred.gif?raw=true'/></td>
  </tr>
<table>
### Cases of multi-view robust



### Ablation study

#### Different combination of loss terms

<img src = 'https://github.com/winterwindwang/Full-coverage-camouflage-adversarial-attack/blob/gh-pages/assets/abaltion_study_loss.png?raw=true'/>

As we can see from the Figure, different loss term plays different role in attacking. For example, the camouflaged car generated by `obj+smooth (we omit the smooth loss, and denotes as obj)` hardly hidden from the detector, while the camouflaged car generated by `iou` successfully suppress the detecting bounding box of the car region, and finally the camouflaged car generated by `cls` successfully make the detector to misclassify the car to anther category.