Machine learning (ML) classifiers can make mistakes that are perceptually and cognitively disturbing to humans. The most notorious examples of such errors are adversarial visual metamers. This paper investigates the phenomenon of adversarial Doppelgängers (AD), which encompasses adversarial visual metamers, and compares the performance and robustness of ML classifiers to human performance.We find that ADs are inputs that are close to each other with respect to a perceptual metric defined in this paper, and show that ADs are qualitatively different from the usual adversarial examples. The vast majority of classifiers are vulnerable to ADs and robustness-accuracy trade-offs may not improve them. Some classification problems do not admit any AD-robust classifiers because the underlying classes are ambiguous. We provide criteria to determine whether a classification problem is well defined; describe the structure and attributes of AD-robust classifiers; introduce and explore the notions of conceptual entropy and regions of conceptual ambiguity for classifiers that are vulnerable to AD attacks; and discuss methods to bound the AD fooling rate of an attack. We define the notion of classifiers that exhibit hypersensitive behavior, that is, classifiers whose only mistakes are adversarial Doppelgängers. Improving the AD robustness of hypersensitive classifiers is equivalent to improving accuracy. We identify conditions guaranteeing that all classifiers with sufficiently high accuracy are hypersensitive.

The analysis of videos taken by active operators recording human interactions and activities in the field presents a new set of challenges. For brevity in this paper we will call such subject centric field grade videos ad hoc videos of events. Human test subjects readily segment ad hoc videos of events into scene-like segments. These segmentations can not be replicated by the state of the art automatic video segmentation algorithms. We propose and evaluate a method to segment ad hoc videos of events into atomic semantics units. Motivated by [Bel74] we call these units sub-scenes. Our experiments show that the segments detected by human subjects are sequences of sub-scenes. Thus the sub-scenes appear to be a semantic version of the video shots that are used to piece together scenes by state of the art video segmentation algorithms.

Numerous applications processing 3D point data will gain from the ability to estimate reliably normals and differential geometric properties. Normal estimates are notoriously noisy, the errors propagate and may lead to flawed, inaccurate, and inconsistent curvature estimates. Frankot-Chellappa introduced the use of integrability constraints in normal estimation. Their approach deals with graphs z = f(x,y)- We present a newly discovered general orientability constraint (GOC) for 3D point clouds sampled from general surfaces, not just graphs. It provides a tool to quantify the confidence in the estimation of normals, topology, and geometry from a point cloud. Furthermore, similarly to the Frankot-Chellappa constraint, the GOC can be used directly to extract the topology and the geometry of the manifolds underlying 3D point clouds. As an illustration we describe an automatic Cloud-to-Geometry pipeline which exploits the GOC.

We present new theoretical results which have implications in answering one of the fundamental questions in computer vision: recognition of surfaces and surface shapes. We state the conditions under which: (i) a surface can be recovered, uniquely, from the tangent plane map, in particular from the Gauss map; (ii) a surface shape can be recovered from the metric and the deforming forces. In case where such conditions are not satisfied we classify all exceptions, i.e. the surfaces and surface shapes for which the recovery and registration problems are ill-posed.