Monitored Distillation for Positive Congruent Depth Completion

We propose a method to infer a dense depth map from a single image, its calibration, and the associated sparse point cloud. In order to leverage existing models that produce putative depth maps (teacher models), we propose an adaptive knowledge distillation approach that yields a positive congruent training process, where a student model avoids learning the error modes of the teachers. We consider the scenario of a blind ensemble where we do not have access to ground truth for model selection nor training. The crux of our method, termed Monitored Distillation, lies in a validation criterion that allows us to learn from teachers by choosing predictions that best minimize the photometric reprojection error for a given image. The result of which is a distilled depth map and a confidence map, or "monitor", for how well a prediction from a particular teacher fits the observed image. The monitor adaptively weights the distilled depth where, if all of the teachers exhibit high residuals, the standard unsupervised image reconstruction loss takes over as the supervisory signal. On indoor scenes (VOID), we outperform blind ensembling baselines by 13.3% and unsupervised methods by 20.3%; we boast a 79% model size reduction while maintaining comparable performance to the best supervised method. For outdoors (KITTI), we tie for 5th overall on the benchmark despite not using ground truth.

Stereoscopic Universal Perturbations across Different Architectures and Datasets

We study the effect of adversarial perturbations of images on deep stereo matching networks for the disparity estimation task. We present a method to craft a single set of perturbations that, when added to any stereo image pair in a dataset, can fool a stereo network to significantly alter the perceived scene geometry. Our perturbation images are "universal" in that they not only corrupt estimates of the network on the dataset they are optimized for, but also generalize to stereo networks with different architectures across different datasets. We evaluate our approach on multiple public benchmark datasets and show that our perturbations can increase D1-error (akin to fooling rate) of state-of-the-art stereo networks from 1% to as much as 87%. We investigate the effect of perturbations on the estimated scene geometry and identify object classes that are most vulnerable. Our analysis on the activations of registered points between left and right images led us to find that certain architectural components, i.e. deformable convolution and explicit matching, can increase robustness against adversaries. We demonstrate that by simply designing networks with such components, one can reduce the effect of adversaries by up to 60.5%, which rivals the robustness of networks fine-tuned with costly adversarial data augmentation.