SCIEN Talk

Virtual reality is a new medium that provides unprecedented user experiences. Eventually, VR/AR systems will redefine communication, entertainment, education, collaborative work, simulation, training, telesurgery, and basic vision research. In all of these applications, the primary interface between the user and the digital world is the near-eye display. While today's VR systems struggle to provide natural and comfortable viewing experiences, next-generation computational near-eye displays have the potential to provide visual experiences that are better than the real world. In this talk, we explore the frontiers of VR/AR systems engineering and discuss next-generation near-eye display technology, including gaze-contingent focus, light field displays, monovision, holographic near-eye displays, and accommodation-invariant near-eye displays.

This talk argues for combining the fields of robotic vision and computational imaging. Both consider the joint design of hardware and algorithms, but with dramatically different approaches and results. Roboticists seldom design their own cameras, and computational imaging seldom considers performance in terms of autonomous decision-making.The union of these fields considers whole-system design from optics to decisions. This yields impactful sensors offering greater autonomy and robustness, especially in challenging imaging conditions. Motivating examples are drawn from autonomous ground and underwater robotics, and the talk concludes with recent advances in the design and evaluation of novel cameras for robotics applications.

Kai Stroeder, Managing Director at Carl Zeiss Smart Optics GmbH, will talk about the Carl Zeiss Smart Glasses.

This will be an informal session with an introduction and prototype demo of the Smart Glasses and an open discussion about future directions and applications.

High-quality 3D object recognition is an important component of many vision and robotics systems. We tackle the object recognition problem using two data representations: Volumetric representation, where the 3D object is discretized spatially as binary voxels - 1 if the voxel is occupied and 0 otherwise. Pixel representation where the 3D object is represented as a set of projected 2D pixel images. At the time of submission, we obtained leading results on the Princeton ModelNet challenge. Some of the best deep learning architectures for classifying 3D CAD models use Convolutional Neural Networks (CNNs) on pixel representation, as seen on the ModelNet leaderboard. Diverging from this trend, we combine both the above representations and exploit them to learn new features. This approach yields a significantly better classifier than using either of the representations in isolation. To do this, we introduce new Volumetric CNN (V-CNN) architectures.

Polarization interferometers are interferometers that utilize birefringent crystals in order to generate an optical path delay between two polarizations of light. In this talk I will describe how I have employed polarization interferometry to make two kinds of Fourier imaging spectrometers; in one case, by temporally scanning the optical path delay with a liquid crystal cell, and in the other, utilizing relative motion between scene and detector to spatially scan the optical path delay through a position-dependent wave plate.

Modern systems-on-a-chip (SoC) have many different types of processors that could be used in computational imaging. Unfortunately, they all have different programming models, and are thus difficult to optimize as a system. In this talk we discuss various standards (OpenCL, OpenVX) and domain-specific programming languages (Halide, Proximal) that make it easier to accelerate processing for computational imaging.

Deep learning has driven huge progress in visual object recognition in the last five years, but this is one aspect of its application to imaging. This talk will provided a brief overview deep learning and artificial neural networks in computer vision, before delving into wide range of application Google has pursued in this area. Topics will include image summarization, image augmentation, artistic style transfer, and medical diagnostics.

In my childhood I invented a new kind of lock-in amplifier and used it as the basis for the world's first wearable augmented reality computer (http://wearcam.org/par). This allowed me to see radio waves, sound waves, and electrical signals inside the human body, all aligned perfectly with the physical space in which they were present. I built this equipment into special electric eyeglasses that automatically adjusted their convergence and focus to match their surroundings. By shearing the spacetime continuum one sees a stroboscopic vision in coordinates in which the speed of light, sound, or wave propagation is exactly zero (http://wearcam.org/kineveillance.pdf), or slowed down, making these signals visible to radio engineers, sound engineers, neurosurgeons, and the like. See the attached picture of a violin attached to the desk in my office at Meta, where we're creating the future of computing based on Human-in-the-Loop Intelligence (https://en.wikipedia.org/wiki/Humanistic_intelligence).

More Information: http://weartech.com/bio.htm

Imaging has become an essential part of how we communicate with each other, how autonomous agents sense the world and act independently, and how we research chemical reactions and biological processes. Today's imaging and computer vision systems, however, often fail in critical scenarios, for example in low light or in fog. This is due to ambiguity in the captured images, introduced partly by imperfect capture systems, such as cellphone optics and sensors, and partly present in the signal before measuring, such as photon shot noise. This ambiguity makes imaging with conventional cameras challenging, e.g. low-light cellphone imaging, and it makes high-level computer vision tasks difficult, such as scene segmentation and understanding.

In this talk, I will present several examples of algorithms that computationally resolve this ambiguity and make sensing and vision systems robust. These methods rely on three key ingredients: accurate probabilistic forward models, learned priors, and efficient large-scale optimization methods. In particular, I will show how to achieve better low-light imaging using cell-phones (beating Google's HDR+), and how to classify images at 3 lux (substantially outperforming very deep convolutional networks, such as the Inception-v4 architecture). Using a similar methodology, I will discuss ways to miniaturize existing camera systems by designing ultra-thin, focus-tunable diffractive optics. Finally, I will present new exotic imaging modalities which enable new applications at the forefront of vision and imaging, such as seeing through scattering media and imaging objects outside direct line of sight.

This workshop will bring together scientists and engineers who are advancing sensor technologies, computer vision, machine learning, head-mounted displays and our understanding of human vision, and developers who are creating new and novel applications for augmented and mixed reality in retail, education, science and medicine.