Visual Computing Seminar (Spring 2018)
The Visual computing seminar is a weekly seminar series on topics in Visual Computing.
Why: The motivation for creating this seminar is that EPFL has a critical mass of people who are working on subtly related topics in computational photography, computer graphics, geometry processing, human–computer interaction, computer vision and signal processing. Having a weekly point of interaction will provide exposure to interesting work in this area and increase awareness of our shared interests and other commonalities like the use of similar computational tools — think of this as the visual computing edition of the “Know thy neighbor” seminar series.
Who: The target audience are faculty, students and postdocs in the visual computing disciplines, but the seminar is open to anyone and guests are welcomed. There is no need to formally enroll in a course. The format is very flexible and will include 45 minute talks with Q&A, talks by external visitors, as well as shorter presentations. In particular, the seminar is also intended as a way for students to obtain feedback on shorter ~20min talks preceding a presentation at a conference. If you are a student or postdoc in one of the visual computing disciplines, you’ll probably receive email from me soon on scheduling a presentation.
Where and when: every Wednesday in BC03 (next to the ground floor atrium). Food is served at 11:50, and the actual talk starts at 12:15.
How to be notified: If you want to be kept up to date with announcements, please send me an email and I’ll put you on the list. If you are working in LCAV, CVLAB, IVRL, LGG, LSP, IIG, CHILI, LDM or RGL, you are automatically subscribed to future announcements, so there is nothing you need to do.
You may add the seminar events to Google Calendar (click the '+' button in the bottom-right corner), or download the iCal file.
Title: Science Art Synergy - the CAD/CAM way
Abstract: The art work of M.C. Escher needs no introduction. We have all learned
In this talk, I will discuss several artistically-related applications
P.S. This talk imposes no prerequisites beyond an open mind.
Title: Beyond the Pixel-Wise Loss for Topology-Aware Delineation
Abstract: Delineation of curvilinear structures is an important problem in Computer Vision with multiple practical applications. With the advent of Deep Learning, many current approaches on automatic delineation have focused on finding more powerful deep architectures, but have continued using the habitual pixel-wise losses such as binary cross-entropy. In this paper we claim that pixel-wise losses alone are unsuitable for this problem because of their inability to reflect the topological impact of mistakes in the final prediction. We propose a new loss term that is aware of the higher-order topological features of linear structures. We also introduce a refinement pipeline that iteratively applies the same model over the previous delineation to refine the predictions at each step while keeping the number of parameters and the complexity of the model constant.
When combined with the standard pixel-wise loss, both our new loss term and our iterative refinement boost the quality of the predicted delineations, in some cases almost doubling the accuracy as compared to the same classifier trained with the binary cross-entropy alone. We show that our approach outperforms state-of-the-art methods on a wide range of data, from microscopy to aerial images.
Title: New tools for localisation and mapping
Abstract: User positioning, self-driving cars and other autonomous robots are all examples of the ubiquity of localisation and mapping. It is well studied in the forms of angulation and lateration—for known landmark positions—and, more generally, simultaneous localisation and mapping (SLAM) and structure from motion (SfM)—for unknown landmark positions. In this talk, I will review existing approaches and discuss some novel priors for these problems. In particular, I will place an emphasis on geometrical techniques using Euclidean distance matrices (EDM)s and discuss extensions to better fit practical applications. In addition, I will attempt to abstract the key essence of localisation and mapping problems, which will allow us to derive fundamental performance limits.
Title: Reconstructing and Perceiving Humans in Motion
Abstract: For man-machine interaction it is crucial to develop models of humans that look and move indistinguishably from real humans. Such virtual humans will be key for application areas such as computer vision, medicine and psychology, virtual and augmented reality and special effects in movies.
Currently, digital models typically lack realistic soft tissue and clothing or require time-consuming manual editing of physical simulation parameters. Our hypothesis is that better and more realistic models of humans and clothing can be learned directly by capturing real people using 4D scans, images, and depth and inertial sensors. Combining statistical machine learning techniques and geometric optimization, we create realistic models from the captured data.
We then leverage the learned digital models to extract information out of incomplete and noisy sensor data coming from monocular video, depth or a small number of IMUs.
I will give an overview of a selection of projects where the goal is to build realistic models of human pose, shape, soft-tissue and clothing. I will also present some of our recent work on 3D reconstruction of people models from monocular video, and real-time joint reconstruction of surface geometry and human body shape from depth data. I will conclude the talk outlining the next challenges in building digital humans and perceiving them from sensory data.
Bio: Gerard Pons-Moll obtained his degree in superior Telecommunications Engineering from the Technical University of Catalonia (UPC) in 2008. From 2007 to 2008 he was at Northeastern University in Boston USA with a fellowship from the Vodafone foundation conducting research on medical image analysis. He received his Ph.D. degree (with distinction) from the Leibniz University of Hannover in 2014. In 2012 he was a visiting researcher at the vision group at the University of Toronto. In 2012 he also worked as intern at the computer vision group at Microsoft Research Cambridge. From 11/2013 until 09/2015 he was a postdoc and later from 10/2015-08/2017 Research Scientist at Perceiving Systems, Max Planck for Intelligent Systems. Since 09/2017 he is heading the group Real Virtual Humans at the Max Planck Institute for Informatics.
His work has been published at the major computer vision and computer graphics conferences and journals including Siggraph, Siggraph Asia, CVPR, ICCV, BMVC(Best Paper), Eurographics(Best Paper), IJCV and TPAMI. He serves regularly as a reviewer for TPAMI, IJCV, Siggraph, Siggraph Asia, CVPR, ICCV, ECCV, ACCV, SCA, ICML and others. He co-organized 1 workshop and 3 tutorials: 1 tutorial at ICCV 2011 on Looking at People: Model Based Pose Estimation, and 2 tutorials at ICCV 2015 and Siggraph 2016 on Modeling Human Bodies in Motion, and the workshop PeopleCap at ICCV'17.
Title: Coherent Hamiltonian Monte Carlo for Light Transport
Abstract: Rendering realistic images involves estimating a difficult and high-dimensional integral. Most classical algorithms approximate this integral with a Monte Carlo estimate over all paths that light can take within a given scene (Path Space). Obtaining a converged, noise-free image thus requires large amounts of computing power.
Modern CPUs have vectorization units which allow carrying out operations simultaneously on 8, 16 or more operands at once. Unfortunately, current state-of-the-art rendering algorithms feature highly incoherent workloads that barely benefit from vectorization. We will give intuition about what “coherence” means in this context, and why it is performance critical.
The talk will include a review of the relevant background, including vector instructions, Physically Based Rendering, MCMC, and Hamiltonian Monte Carlo.
Title: An Interlocking Method for 3D Assembly Design and Fabrication
Abstract: 3D assemblies such as furniture involve multiple component parts. Rather than relying on additional fasteners such as nails and screws to connect the parts, component parts can be interlocked into a steady assembly based on their own geometric arrangements. This talk revisits the notion of 3D interlocking, and explores the governing mechanics of general 3D interlocking assemblies. From this, constructive approaches are developed for computational design of various new interlocking assemblies such as puzzles, 3D printed objects and furniture. These interlocking assemblies are ready for fabrication and their steadiness have been validated in our experiments.
Title: Looking beyond pixels: theory, algorithms and applications of continuous sparse recovery
Abstract: Sparse recovery is a powerful tool that plays a central role in many applications. Conventional approaches usually resort to discretization, where the sparse signals are estimated on a pre-defined grid. However, the sparse signals do not line up conveniently on any grid in reality.
Title: Learning Light-Transport using Variational Auto Encoders
Abstract: In the real world, many objects are made out of materials which allow some light to be transmitted and scattered internally. This effect is called subsurface scattering and can have a large impact on the visual appearance of an object. Human skin, milk, marble and many other materials exhibit subsurface scattering. When rendering virtual scenes containing such objects, we have to simulate these internal light bounces, which is an expensive process.
In the past, different algorithms have been proposed to approximate internal scattering by simpler, analytical models. However, these approximations are inaccurate in many scenarios, since they often do not account for the scene geometry.
In this ongoing project, we try to accelerate rendering of subsurface scattering using machine learning. Instead of exhaustively simulating internal light scattering, we use a Variational Auto-Encoder (VAE) to estimate light transport more efficiently. Conditioned on the geometry of the scene, the VAE produces samples from the distribution of light scattered inside an object. Our method is readily embedded in conventional rendering systems.
In this talk, we present the necessary background in subsurface scattering rendering as well as variational auto-encoders, and no prior knowledge in either topic is assumed.
Title: Representation Learning for semi-supervised 3D Human Pose Estimation
Abstract: Modern 3D human pose estimation techniques rely on deep networks, which require large amounts of training data. I'll present two semi-supervised approaches to overcome this problem.
Title: Robust Elastic Metamaterial Design for Additive Fabrication
Abstract: 3D printing can efficiently manufacture geometry of arbitrary complexity but offers only limited control over the elastic material properties of the printed part. In this talk, I will present my work on designing elastic metamaterials for 3D printing: periodic microstructures that are tuned to emulate a large space of elastic materials. Since microstructures typically contain thin features that concentrate stress, they are prone to plastic deformation and fracture even under mild deformations. A key goal of my work is to design microstructures that minimize these stress concentrations, improving the metamaterials’ robustness in practice. The design algorithm I developed is based on a efficient, exact solution to the worst-case stress analysis problem for periodic microstructures which supports several failure criteria (e.g., maximum principal stress or von Mises stress); it produces microstructures that minimize worst-case stresses while achieving a particular target elasticity tensor and satisfying fabrication constraints. I also will present a design tool to optimally apply these metamaterials to achieve high-level deformation goals.
Title: The Layer Laboratory: A Calculus for Additive and Subtractive Composition of Anisotropic Surface Reflectance
Abstract: We present a versatile computational framework for modeling the reflective and transmissive properties of arbitrarily layered anisotropic material structures. Given a set of input layers, our model synthesizes an effective BSDF (bidirectional scattering distribution function) of the entire structure, which accounts for all orders of internal scattering and is efficient to sample and evaluate in modern rendering systems.
Title: Learning to Mean-Shift in O(1) for Bayesian Image Restoration
Abstract: Finding strong oracle priors is an important topic in image restoration. In this talk, I will show how denoising autoencoders (DAEs) learn to mean-shift in O(1), and how we leverage this to employ DAEs as generic priors for image restoration. I will also discuss the case of Gaussian DAEs in a Bayesian framework, where the degradation noise and/or blur kernel are unknown. Experimental results demonstrate state of the art performance of the proposed DAE priors.