Title: Data-driven Geometry Processing
Many geometry processing tasks require understanding geometric data from human perspective. Recent growth of large online repositories of 3D shapes and advances in machine learning enable building data models that can be used to understand similarities, variations, semantics, and functionality of 3D objects. There two main challenges in using machine learning for geometry processing. First, we need to collect human annotations to learn functional and semantic attributes of shapes. And second, we need to develop novel geometric representations that are compatible with state-of-the-art machine learning algorithms. To address the first challenge, we developed several techniques that significantly reduce the cost of human supervision by combining several types of crowd-sourced worker tasks, automatic label propagation, and meta-data that comes for free with the 3D models. In the second part of my talk I will discuss the challenges, existing solutions, and open research problems in geometric representations for machine learning.
Title: Towards the Expressive Design of Virtual Worlds : Combining Knowledge and Control
Despite our great expressive skills, we humans lack an easy way of conveying the 3D worlds we imagine. While impressive advances were made in the last fifteen years to evolve digital modeling systems into gesture-based interfaces enabling to sketch or sculpt in 3D, modeling is generally limited to the design of isolated, static shapes. In contrast, virtual worlds are composed of distributions or assemblies of elements which are too numerous to be created or even positioned one by one; the shapes of these elements may heavily depend on physical laws and on the interaction with their surroundings; and many of them may be animated, meaning that they should not only be designed in space, but also over time. In this talk, I will explore the recent extensions of expressive modeling metaphors such as sketching, painting, transfer and sculpting, to such complex cases. I will show that models for shape and motion need to be redefined from a user-centered perspective, and in particular embed the necessary knowledge to make them respond in an intuitive way to the user’s control gestures.
PhD Thesis Proposal Examination: 3D Modeling of Urban Scene Using Quadrotors - Rui Huang
PhD Depth Examination: Camera Pose Estimation in Structure-From-Motion - Zhaopeng Cui
Structure-from-motion (SfM) is a fundamental problem in computer vision. It refers to the process of estimating 3d scene structures and camera poses simultaneously from multiple 2d images. Conventional SfM systems often consist of three steps:
1) estimation of point correspondences and relative poses between images
2) estimation of camera motions and 3d points, and
Dr. Richard Zhang will be presenting the talk "Why is Computer Graphics Hard?" as part of SFU's School of Computing Science Colloquium series of research talks by faculty members and grad students.
Title : From Professional Tools to Consumer Fun
The automatic creation of man-made 3D objects is an active area in computer graphics. Mixing and blending of components or sub components from existing example shapes can help users quickly produce interesting and creative designs. A key factor for automating this task is using computer algorithms that can map between objects of different shape and structure. However, due to the coarse correspondence computed by current matching algorithms, automatic shape blending is mainly limited to the substitution of compatible part sets.
Title: Comparing geometric representations for real-time graphics
These days, on the internet, you may have read polarized opinions such as “polygons are the only serious representation in any serious 3D software”, or “this is the end for polygons”, when polygons are compared to one of the other geometric representations. While both opinions are ridiculous, we will review some of the advantages and disadvantages of polygons, point cloud and voxels as geometric representations in real-time 3D applications, to get a better appreciation for why those polarized opinions exist, and why, in the end, those opinions do not really matter.
Food, activities and fun for GrUVi members, family and friends.
Location: Barnet Marine Park
We introduce an unsupervised analysis of both homogeneous and heterogeneous shape collections, aiming at organizing shapes based on their similarity in structure. We derive the idea of graph representation of shape structure from previous works and a novel graph editing distance based structure matching cost is defined. For any pair of shapes, we propose a searching scheme to find the best matching pair of graphs with the minimal cost.
Space-saving, or collapsible, objects are ubiquitous in our living and working space. They can adjust configurations to either perform their intended functionality or save space, for example, while storing and shipping. This additional space-saving characteristic of collapsible objects comparing to their non-collapsible counterparts makes them more preferable, especially in environments where space is costly.
The majority of existing methods for delineating trees from LIDAR point cloud use a region growing approach. Seed points representing the highest point of the trees (tree-peaks) are detected in the point cloud. The remaining points are iteratively assigned to one of the seed points, thus growing the region representing trees. The tree-peak detection methodologies are based on local geometry analysis, identifying locally highest points within some appropriately sized neighborhood as tree-peaks.
Title: State of the Art in Surface Reconstruction from Point Clouds
The area of surface reconstruction has seen substantial progress in the past two decades. The traditional problem addressed by surface reconstruction is to recover the digital representation of a physical shape that has been scanned, where the scanned data contains a wide variety of defects.
Mesh traversal is a common and essential geometry processing problem in computer graphics. The traversal typically processes each face in a mesh in a systematic and consistent order for different applications such as mesh compression, rendering and mesh simplification. While cache-efficient mesh traversal methods where data and computations are reordered for good cache reuse have been well-studied, their performances are limited by implicit(automatic) memory management. In this work we explore optimizations on Explicitly Managed Memory (EMM) systems.
In the proposed thesis, we address two challenges related to skeletons. The first is to formulate novel ways to define and compute curve skeletons - a specific type of skeleton. The second is to employ skeletons to enhance surface reconstruction for geometry affected by severe amounts of missing data. In solving these challenges we discuss three different approaches.
Title: Discovering Similarities In Diverse Collections of 3D Shapes
Title: Simulating Human Locomotion: Optimization, Uncertainty, and Biomechanics
Locomotion, specifically walking and running, are common and essential human movements. The ability to create physically and biomechanically plausible simulations of locomotion is of interest to applications ranging from game content creation to pathological gait analysis, and can contribute to our understanding of motor control.
However, while humans can move on varied terrains, start and stop on a dime, and recover from trips with ease, getting simulated humanoids to simply walk forward without falling is a challenging task. Achieving locomotion requires solving a high-dimensional, nonlinear, and underactuated control problem. Furthermore, out of all the control strategies that accomplish the task, how do we select one that produces human-like movements? In this talk, the speaker will present an approach to simulate and control 3D humanoid locomotion that produces results matching human data to a much greater extent than previous state-of-the-art.
Title: Towards Automatic Visual Content Creation
Computer graphics has been very successful. However, an important problem still remains unsolved. High quality graphical content (such as 3D models and realistic images or animations) is difficult to create, which limits graphics to mass market products, such as games and movies. To make graphics a media for our daily communication, we must bridge the gap between ordinary people and graphics professionals, and make visual content easy to create for the general public. This talk introduces the speaker's recent work towards this goal.
Physics simulations are widely recognized to be crucial tools for complex special effects in feature films, and real-time simulations are often central game-play elements in modern computer games. There are, however, inherent difficulties with these simulations: we are still very far from being able to accurately simulate the complexity of nature around us. Additionally, the numerical methods that are commonly used are notoriously difficult to fine-tune and control. The central goal of the speaker's research is to address these issues with novel multi-physics solvers.