SCIEN Talk

3D Computer Vision (3D Vision) techniques have been the key solutions to various scene perception problems such as depth from image(s), camera/object pose estimation, localization and 3D reconstruction of a scene. These solutions are the major part of many AI applications including AR/VR, autonomous driving and robotics. In this talk, I will first review several categories of 3D Vision problems and their challenges. Given the category of static scene perception, I will introduce several learning-based depth estimation methods such as PlaneRCNN, Neural RGBD, and camera pose estimation methods including MapNet as well as few registration algorithms deployed in NVIDIA's products. I will then introduce more challenging real world scenarios where scenes contain non-stationary rigid changes, non-rigid motions, or varying appearance due to the reflectance and lighting changes, which can cause scene reconstruction to fail due to the view dependent properties. I will discuss several solutions to these problems and conclude by summarizing the future directions for 3D Vision research that are being conducted by NVIDIA's learning and perception research (LPR) team.

Cancer is a surgically treated disease; almost 80% of early stage solid tumors undergo surgery at some point in their treatment course. The biggest gap in quality remains the high rate of tumor-positive margins in surgical resections. The biggest barrier is that only a limited amount of the tissue can be sampled for frozen section analysis (< 5%). The biggest challenge is to develop equipment to direct frozen section analysis to the most area on the specimen most likely to contain a positive margin. To this end, we developed intraoperative devices to leverage molecular imaging during and immediately after cancer resections.

This talk will present the methods and procedures used to produce the first results from the Event Horizon Telescope. It is theorized that a black hole will leave a "shadow" on a background of hot gas. Taking a picture of this black hole shadow could help to address a number of important scientific questions, both on the nature of black holes and the validity of general relativity. Unfortunately, due to its small size, traditional imaging approaches require an Earth-sized radio telescope. In this talk, I discuss techniques we have developed to photograph a black hole using the Event Horizon Telescope, a network of telescopes scattered across the globe. Imaging a black hole's structure with this computational telescope requires us to reconstruct images from sparse measurements, heavily corrupted by atmospheric error.

Scientific researchers now utilize advanced microscopes to collect very large volumes of image data. These volumes often contain morphologically complex structures that can be difficulty to comprehend on a 2D monitor, even with 3D projection. syGlass is a software stack designed specifically for the visualization, exploration, and annotation of very large image volumes in virtual reality. This technology provides crucial advantages to exploring 3D volumetric data by correctly leveraging neurological processes and pipelines in the visual cortex, reducing cognitive load and search times, while increasing insight and annotation accuracy. The talk will provide a brief overview of new microscope technology, a description of the syGlass stack and product, some real use-cases from various labs around the world, and conclude with predictions and plans for the future of scientific communication.

For the vast majority of people in the world, the best camera they have ever owned is in their current cell phone. Sales of phone camera modules approached $30 billion in 2018, almost three times the sales of all lasers, and will soon exceed four times the sales of all lasers. This is one of the most ubiquitous and successful optical devices ever. Fundamental laws of physics limit the performance of smartphone cameras, and these laws act against the marketing-driven aspiration for thinner and thinner camera modules. I shall show that the single most important optical parameter is the lens diameter D.

Mobile photography has been transformed by software. While sensors and lens design have improved over time, the mobile phone industry relies increasingly on software to mitigate physical limits and the constraints imposed by industrial design. In this talk, I'll describe the HDR+ system for burst photography, comprising robust and efficient algorithms for capturing, fusing, and processing multiple images into a single higher-quality result. HDR+ is core imaging technology for Google's Pixel phones - it's used in all camera modes and powers millions of photos per day. I'll give a brief history of HDR+ starting from Google Glass (2013), present key algorithms from the HDR+ system, then describe the new features that enable the recently released Night Sight mode.

Neural networks have surpassed the performance of virtually any traditional computer vision algorithm thanks to their ability to learn priors directly from the data. The common encoder/decoder with skip connections architecture, for instance, has been successfully employed in a number of tasks, from optical flow estimation, to image deblurring, image denoising, and even higher level tasks, such as image-to-image translation.

To improve the results further, one must leverage the constraints of the specific problem at hand, in particular when the domain is fairly well understood, such as the case of computational imaging.

In this talk I will describe a few of my recent projects that build on this observation, ranging from reflection removal, to novel view synthesis, and deblurring.

Disney Research has been actively pushing the state-of-the-art in digitizing humans over the past decade, impacting both academia and industry. In this talk I will give an overview of a selected few projects in this area, from research into production. I will be talking about photogrammetric shape acquisition and dense performance capture for faces, eye and teeth scanning and parameterization, as well as physically based capture and modelling for hair and volumetric tissues.

Precise determination of the position of a single point source (e.g. fluorescent molecule/protein, quantum dot) is at the heart of microscopy methods such as single particle tracking and super-resolution localization microscopy ((F)PALM, STORM). Localizing a point source in all three dimensions, i.e. including depth, poses a significant challenge; the depth of field of a standard high-NA microscope is fundamentally limited, and its pointspread-function (PSF), namely, the shape that a point source creates in the image plane, contains little information about the emitter's depth. Various techniques exist that enable 3D localization, prominent among them being PSF engineering, in which the PSF of a microscope is modified to encode the depth of the source. This is achieved by shaping the wavefront of the light emitted from the sample, using a phase mask in the pupil (Fourier) plane of the microscope.

In this talk, I will describe how our search for the optimal PSF for 3D localization, using tools from estimation theory, led to the development of microscopy systems with unprecedented capabilities in terms of depth of field and spectral discrimination. Such methods enable fast, precise, non-destructive localization in thick samples and in multicolor. Applications of these novel advances will be demonstrated, including super-resolution imaging, tracking biomolecules in living cells and microfluidic flow profiling. I will also present our most recent results: 1. Application of deep learning for solving difficult localization problems (high density, low SNR, multicolor imaging), and 2. Precise refractometry from minute volumes by super-critical-angle fluorescence.

Exploiting to see beyond the diffraction-limit of optical microscopies is of great significance. State-of-the-art solutions of super-resolution microscopy, like STED and STORM approaches, rely on the fluorescent effect of labeling samples. It is still challenging to obtain the super-resolution for unlabeling samples without fluorescent effect. To this end, we have developed a novel super-resolution method, called Spatial Frequency Shift (SFS), to realize the deep super-resolution with or without fluorescent effect in wide field imaging. The principle and the applications of this SFS technique will be presented.