Adversarial 3D Objects Against Monocular Depth Estimators

The state-of-the-art solutions for the reconstruction of the depth from 2D images (monocular depth estimation) are deep neural networks. Although the neural networks in the field of computer vision generally provide good performance, they are quite sensitive to specially crafted inputs, called adversarial examples. It is possible to create these adversarial examples solely by manipulating the physical world with special lighting, painted textures, or 3D shapes. Unlike most of the existing literature in the field of monocular depth estimation, this paper focuses on adversarial 3D shapes instead of painted textures or lighting. According to our results, the adversarial 3D shapes can worsen the performance of monocular depth estimators, achieving an effect size comparable to the differences between subsequent state-of-the-art models. We published our implementation at: https://github.com/mntusr/threedattack1