UMN Visual Computing & AI Seminar
03/20/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Zhihang Deng
Providing 3D natural interaction using bare hands with haptics feedback is difficult but is in demand for many VR applications such as architecture design, medical training, education, etc. Passive haptics provides a compelling haptics experience by using haptics-physical props to which virtual objects are registered. Meanwhile, using self-avatar and passive haptics together can enhance VR's immersed experience, but it is hard to apply due to the difficulties in acquiring hand model, registering virtual hands registration to the real hands, and building an analogous physical prop for providing passive haptics. In this talk, I will talk about how we use Pseudo-haptics, an illusion caused by inconsistency of visual cue and haptics cue, to provide an illusion of object's length. And we also study different types of self-avatar hand representations influence on the length illusion in VR.
Bio: Zhihang is a Ph.D. student. He is working on the Human Perception of Virtual Reality with Professor Victoria Interrante. Currently, he is working on providing haptics feedback using passive haptics technology, pseudo haptics. Specifically, he is working on mapping simple physical props to a various shape of virtual objects in VR and studying the influence of different self-avatar hands representations on pseudo haptics in VR. He also interested in Computer Vision and Computer Graphics.
03/27/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: George Brown
Title: Accurate Dissipative Forces in Optimization Integrators
Bio: George is a third year PhD student advised by Prof. Rahul Narain. His research area is physics-based animation, with a focus on optimization-based methods for fast and efficient numerical simulation.
04/03/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Jae Shin Yoon
Title: 3D Semantic Trajectory Reconstruction from 3D Pixel Continuum
A 3D trajectory representation of human interactions is a viable computational model that measures microscopic actions at high spatial resolution without prior scene assumptions. Unfortunately, the representation is a lack of semantics, which fundamentally prevents from the computational behavioral analysis. It is important to know not only where a 3D point is but also what it means and how associated with other points. In this talk, I will first present how to reconstruct 3D trajectories using multiple cameras system that emulates the 3D pixel continuum. Given 3D trajectories, I will mainly focus on how to optimally associate the semantics from 2D with 3D trajectories. Lastly, I will briefly introduce my ongoing project on dense dynamic reconstruction using multiple cameras system.
Bio: Jae Shin is a first-year Ph.D. student advised by Prof. Hyun Soo Park. Currently, he is working on the Trajectory Reconstruction in the 3D space and its application using multiple cameras system. His research interests include Computer Vision, 3D Vision, and Machine Learning. For more information about him: cs.umn.edu/~jsyoon
03/06/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Prof. Junaed Sattar
Title: The vagaries of robot sea trials: from Matlab to Madness
Field robotics is all about deploying robotic systems in natural, and often hostile, conditions to evaluate their performance in realistic settings. In the case of our Interactive Robotics and Vision Lab, it involves deploying autonomous underwater robots in open-water environments -- open seas and lakes. This talk will try to give some insights into the journey from the drawing board to the dive board, with a focus on highlighting the process of conceiving algorithms for underwater robotics, specifically for visual perception, learning, human-robot interaction, and navigation, to field testing the entire system.
Bio: I'm an assistant professor at the Department of Computer Science and Engineering at the University of Minnesota, and a MnDrive faculty. I am the founding director of the Interactive Robotics and Vision Lab, where we investigate problems in field robotics, robot vision, human-robot communication, assisted driving and applied (deep) machine learning, not to mention developing rugged robotic systems. My graduate degrees are from McGill University in Canada, and I have a BS in Engineering degree from the Bangladesh University of Engineering and Technology. Before coming to the UoM, I worked as a post-doctoral fellow at the University of British Columbia where I worked on service and assistive robotics, and at Clarkson University in upstate New York as an Assistant Professor. Find me at junaedsattar.org, and the IRV Lab at irvlab.cs.umn.edu or @irvlab on Twitter.
02/20/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Bobby Davis
Title: Motion Planning Under Uncertainty
Predictive planning approaches have revolutionized the ability of robots to move through complex environments with dynamic obstacles. This talk will cover our recent work on robotic motion planning under varying sources of uncertainty. Specifically, we'll focus on how we can include knowledge about how uncertainty evolves to improve local planning in two different contexts. The first context will discuss on how uncertainty can be used to guide and improve information acquisition for UAVs and other robots in human environments. The second context will discuss how we can exploit structure in the uncertainty of the predicted future states of other agents to create safer and faster trajectories.
Bio: Bobby is a fifth year Ph.D. student working with Stephen J. Guy. His research focuses on robot motion planning, especially when there is uncertainty in the robot state or environment.
02/13/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Prof. Richard Linares (Aerospace Engineering and Mechanics Department)
Title: Hamiltonian Monte Carlo Approach to Bayesian Inversion, Nonlinear Filtering, and Optimal Control
This presentation investigates the application of Hamiltonian Monte Carlo (HMC) samplers for solving Bayesian inversion and nonlinear filtering problems that arise in many engineering applications. The HMC approach provides an improvement over Metropolis-Hastings based algorithms by reducing the correlation between successive sampled states by using a Hamiltonian dynamical system evolution. Furthermore, the HMC sampler can be formulated as a dynamical system which is ergodic with respect to the target density. Using the theory of HMC samplers the nonlinear filtering problem for continuous dynamics and discrete measurements is formulated using two stochastic differential equations, one equation for the dynamics of the system and one equation for the measurement update. Furthermore, the filtering problem can be stated as the solution of two versions of the Fokker Planck Kolmogorov Equation (FPKF) one in physical time and one using an auxiliary time variable. Therefore, any existing method for solving the uncertainty propagation problem for nonlinear stochastic systems can now be used to also perform the measurement update. The proposed approach is shown to overcome the Degeneracy Phenomenon associated with particle filters. The use of stochastic optimal control theory to improve the performance of such a method is discussed. A few simple examples are presented to highlight the application of this theory to practical problems.
Bio: Professor Linares's research interests are state and parameter estimation, uncertainty quantification theory, and information fusion for Space Situational Awareness. Dr. Linares joined the faculty of University of Minnesota's Department of Aerospace Engineering and Mechanics as a Professor (Assistant) in 2015 after a short tenure as a research associate at the US Naval Observatory. Prior to moving to Washington, D.C., he held a Director's Postdoctoral Fellow position at Los Alamos National Labs. He has co-authored over 45 conference and journal papers in areas related to space situational awareness, reinforcement learning, artificial intelligence, spacecraft systems, and information fusion. Dr. Linares is the recipient of the AFOSR Young Investigator Research Program Award.
02/06/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Michael Tetzlaff
Title: IBRelight: Image-Based 3D Renderering of Color Appearance for Cultural Heritage
Image-based rendering and relighting has been extensively studied by computer graphics researchers, but until now has not been made accessible to the cultural heritage community. IBRelight is a new 3D rendering tool that addresses this disparity. It leverages existing photogrammetry techniques, applied to camera-mounted flash photographs, to estimate the camera poses and the geometry of the object. IBRelight then reprojects and blends the photographs in a way that emulates the desired new lighting configuration, which may consist of point lights, an environment map, or both. The interface for specifying the camera and lighting is designed to be user-friendly and is based on other modern 3D applications.
Bio: Michael is a fifth year Ph.D. student working with Gary Meyer in 3D computer graphics, with a focus on image-based rendering and relighting for cultural heritage applications.
01/30/2018 Tue 2-3pm @ KHKH 4-192A
Speaker: Jung Who Nam
Title: Worlds-in-Wedges: Combining WIM and Portal VR Techniques to Support Comparative Scientific Visualization
We present Worlds-in-Wedges, a virtual reality visualization and 3D user interface technique for making simultaneous visual comparisons of multiple worlds (e.g., spatial 3D datasets, historical reconstructions, other data-driven 3D scenes). Comparison is a crucial task for data analysis, but it is not well understood how to facilitate comparative analysis for datasets best displayed in VR, where the tradition is to become immersed in a single world at a time. Our solution is to construct a visualization where it is possible to be in multiple worlds at once while also maintaining an understanding of how these worlds relate and being able to interact with each world. This is accomplished via a three-level visualization. The first level, worlds-in-context, visualizes the relationship between the different worlds (e.g., a map for worlds that are related in space, a timeline for worlds that are related in time). The second level, worlds-in-miniature, is a multi-instance version of the classic World-in-Miniature VR interface and visualizes the user’s position and orientation within each world. The third level, worlds-in-wedges, divides the virtual space surrounding the user into wedge-shaped volumes, where each wedge acts as a volumetric portal, displaying a portion of one world. This visualization is tightly integrated with a bimanual 3D user interface to control the processes of creating new wedges, adjusting their relative size, navigating through one or multiple worlds, and querying data. The new techniques are demonstrated and evaluated via an application to comparing plots from the US Forest Service’s Forest Inventory and Analysis dataset. An evaluation together with collaborating domain scientists suggests the technique usefully complements traditional analyses of these data and shows promise for use both by scientists and as a public-facing storytelling tool.
Bio: Jung Who is a third year Ph.D. student working with Daniel F. Keefe. His research focuses on interactive data visualization in Virtual Reality and Augmented Reality.
12/12/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Prof. Rahul Narain
Title: Towards Accurate Dissipation in Physics-Based Animation
Computer animation has traditionally not paid a great deal of attention to dissipative forces, probably because the use of unconditionally stable numerical methods (necessary to prevent the animator or player from ever blowing up the simulation) already introduces too much artificial dissipation into the system. However, as the field evolves towards new applications outside entertainment and the methods in use become more accurate, it is becoming necessary to model dissipative forces such as friction with the same level of fidelity.
In this talk, I will share two projects our group has been working on in this direction. First, we are working to accurately compute frictional interactions of cloth with solid objects and with other cloth sheets, by combining our adaptive remeshing strategy with an efficient solver for frictional contact constraints. Second, we are extending the optimization-based formulation of simulation (which underlies our ADMM-based method and other popular techniques) to support general nonlinear dissipative forces without losing speed and accuracy, thus allowing fast, realistic animation of arbitrary soft materials.
12/05/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Prof. Hyun Soo Park
Title: 3D Pixel Continuum
Now cameras are deeply integrated in our daily lives, e.g., Amazon Cloud Cam and Nest Cam, reaching soon towards 3D pixel continuum---every 3D point in our space is observed in a form of multiple view pixels by a network of ubiquitous cameras. Such cameras open up a unique opportunity to quantitatively analyze our detailed interactions with scenes, objects, and people continuously, which will facilitate behavioral monitoring for the elderly, human-robot collaboration, and social tele-presence. In this talk, I will introduce our multicamera system built in Shepherd Laboratory that can emulate the 3D pixel continuum that consists of 69 HD synchronized cameras. This is still on-going work but I would like to share our direction and effort towards 3D behavioral modeling.
Specifically, I will address the following questions:
- Why do we need many cameras?
- What are the hardware challenges building the 3D pixel continuum?
- How accurately can we measure our behaviors?
- How can existing vision based systems benefit from the 3D pixel continuum?
- What will be the future of the continuum?
11/28/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Cheng Peng
Title: View selection for optimal reconstruction
I focus on 3D reconstruction using cameras and laser sensors. In the orchard environment, in order to manage the crops more efficiently, obtaining the 3D model of the orchard can be beneficial to the grower to evaluate multiple important traits of the plants such as the yield information, tree height and volume, etc.
My talk will focuses on two parts.
I first started from using both the LIDAR and camera sensors to reconstruct the orchard. Using the idea of super pixels, we are able to find accurate association between image features (SIFT) and LIDAR points and reconstruct the orchard accurately.
Secondly, aside from the side views of the orchard, we started looking at the orchard from the top. Although structure from motion has been well studied, most of the algorithms rely on the heuristics to maintain high quality results. The top view presents a cleaner geometry, which enables us to approach the reconstruction problem from a geometrical point of view.To improve the reconstruction quality, we formulated our problem based on a more direct evaluation. By modeling feature uncertainty as a cone, we can find the uncertainty of any target in metric distance. Thus minimizing the metric distance immediately improves the reconstruction quality.
We presented the theoretical proof and an algorithm that helps determine the best subset of views to reconstruct the desired region optimally, where optimal is defined over the minimum metric error to the true locations of each point in the reconstruction.
Bio: Cheng is a PhD student advised by Prof. Volkan Isler. His research area is mainly on real-time 3D Reconstruction and Mapping, with a current focus on geometrical representation of agricultural environment.
11/21/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Zahra Forootaninia
Title: Uncertainty Models for TTC-Based Collision Avoidance
Abstract: Navigating multiple agents through an environment without any collisions is one of the challenges for robot development. Agents have to detect and avoid collision among several moving characters locally without having a knowledge of the entire environment. Therefore, local collision avoidance models play an important role in multi-agent navigation and planning. In this work, we tackle the problem of uncertainty in the sensing data for multi-agent navigation based on a collision avoidance model called Time-To-Collision (TTC). We propose two ways of assuming uncertainty in agent's path; the isotopic model that considers all possible uncertainties and the adversarial model has the uncertainty only in the direction of a head-on collision. We analyze our methods mathematically and experimentally to show that these two models can produce collision-free interaction between agents.
Bio: I am a Ph.D. student working with Prof. Rahul Narain on physics-based animation techniques. Currently, my focus is on fast and efficient numerical methods for animating crowd and fluid.
11/07/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Daniel Orban
Title: Interactive Visual Querying of Large Parameter Spaces: Shooting Things Accurately while Fluidly Controlling Stress Under Pressure
Abstract: Scientists love exploring the new frontiers of the universe. Engineers enjoy searching for solutions to extremely difficult problems. Unfortunately, due to the large, multidimensional, and nonlinear nature of their respective domains, it is extremely easy to get lost in the data. For example, a medical device designer might want to optimize how a device affects a patient’s specific anatomy, but there are many possible solutions and each solution’s data is large and complex, making searching and simultaneous visual comparison extremely hard. This talk asks the question, “How can we interactively explore a large sparse parameterized space where each data instance is also large?” To work towards a solution, I discuss the combination of two recent projects: Quest and Bento Box. Quest is an application in shock physics that allows scientists to discover which next experiments should be run in order to optimize the knowledge of a high-dimensional sparse continuous system. It uses regression and interpolation techniques on top of a large data ensemble to interactively predict results and estimate uncertainty. Quest also gives the scientist the ability to influence decisions via directly manipulating visually encoded parameters. Bento Box approaches this exploration problem from the other side, assuming each data instance is large and complex. It uses spatial and temporal sampling to visualize and compare specific user-defined features of interest. We use Bento Box to explore the complex fluid-structure interaction of cardiac leads in the right atrium of the heart. I argue that the methods used in Quest and Bento Box can aid in our approach to interactively interpolate and comparatively explore a large number of instances where each instance is also large.
Bio: Dan is a third year PhD student advised by Prof. Dan Keefe. His research focuses on interactive large-scale data visualization and visual parameter space analysis.
10/31/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Wenbo Dong
Title: 3D Computer Vision in Orchard Environments
Established 3D computer vision techniques often fail to perform well in orchard-like environments. In this talk, I will introduce three examples of our work on 3D computer vision in such environments. First, I will show how to accurately calibrate a 2D Laser-Rangefinder (LRF) with a camera such that we can build a colored 3D map of the orchard using the 2D-camera rig. Second, I will present how UAVs can obtain linear velocities using computer vision in order to navigate through an orchard at a low attitude. Last, I will briefly introduce my ongoing work on 3D reconstruction of orchard rows and extracting semantic information such as tree trunk diameter.
Bio: Wenbo is a PhD student advised by Prof. Volkan Isler. His research area is mainly on 3D computer vision, with a current focus on semantic reconstruction of agricultural environment.
10/24/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Matthew Overby
Title: ADMM ⊇ Projective Dynamics: Fast Simulation of Hyperelastic Models with Dynamic Constraints
Elastic deformation is an essential component of cinema and visual effects. Modern simulation techniques allow the creation of real and fantasy characters with a level of realism that would otherwise be impossible. We apply the alternating direction method of multipliers (ADMM) optimization algorithm to implicit time integration of elastic bodies. As ADMM is a general purpose optimization algorithm applicable to a broad range of objective functions, it permits the use of nonlinear constitutive models and hard constraints while maintaining a high level of speed, parallelizability, and robustness. We further extend the algorithm to improve the handling of dynamically changing constraints such as skin sliding and contact.
Bio: Matt is a third year PhD student advised by Dr. Rahul Narain. His research area is physics based animation, with a focus on elastic deformation.
10/10/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Prof. Dan Keefe
Title: Experiential Analytics: From Large-Scale Public Art to Immersive Visualization — When to Walk Inside Your Data
The technical barriers to immersion are all but gone today, but the question remains — is it a good idea to immersive yourself in data? I mean quite literally, does it ever actually make sense to stand inside a large-scale virtual (or physical) data visualization and walk through your data? In this short talk, I'll reflect on several recent immersive visualization projects by my group and our collaborators, some immersive virtual reality data visualizations as well as two large-scale public art installations. In each case, the special ingredient that makes each work as an example of immersive analytics is the need to not only analyze the data but also to experience it via a first-person perspective. Thus, I wonder if we should be calling immersive analytics something more like experiential analytics. How do we understand and design for this experience? How do we quantify it? When is it essential and for whom? Are there counter examples where it is unnecessary and slows down the analytical process? I hope this talk will inspire reflection and discussion on these topics and the exciting future of immersive analytics, as clearly evidenced by this year's workshop.
10/03/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Shan Su
Title: Planning from First-person View
First-person videos record human activities from camera wearer’s own perspective. It captures subtle social and physical interactions of the wearer as following her/his visual attention. In this talk, I will discuss how such first-person perception can be applied to predicting human behaviors and planning motion. First, I will present predicting the collective movement of basketball players from their first-person videos. We leverage two visual cues embedded in the first-person videos: visual semantics of spatial and social layout around the person, and joint attention that link the individuals to a group. Second, I will present creating a visual scene from a first person action. We introduce a concept called ActionTunnel: a 3D virtual tunnel that encodes the wearer's visual experience while moving into the scene. Such abstraction allows us to associate distinctive images w.r.t the afforded first person action.
09/19/2017 Tue 2-3pm @ KHKH 4-192A
Speaker: Prof. Volkan Isler
Title: Robotic Data Gathering in Agricultural and Environmental Monitoring
In this talk, I will give an overview of our efforts to build teams of autonomous aerial, ground and/or surface vehicles for data gathering. After a general overview, I will present an algorithm for collecting bearing data and analyze its performance.
For more information about our work: http://rsn.cs.umn.edu/