Adversarial examples have exposed the weakness of deep networks. Careful modification of the input fools the network completely. Little work has been done to expose the weakness of handcrafted features in adversarial settings. In this work, we propose novel adversarial perturbations for handcrafted features. Pixel level analysis of handcrafted features reveals simple modifications which considerably degrade their performance. These perturbations generalize over different features, viewpoint and illumination changes. We demonstrate successful attack on several well known pipelines (SLAM, visual odometry, SfM etc.). Extensive evaluation is presented on multiple public benchmarks.
Supplementary material (PDF)