CAREER: Deep Learning Empowered Nonlinear Deformable Model

Clemson University, Clemson SC

Investigators

Abstract

Everything in the world deforms, so modeling high-quality deformations becomes a core algorithmic ingredient for serious and realism-driven visual applications such as high-fidelity animation, virtual reality, medical data analysis, surgical simulation, and digital fabrication/prototyping, to name just a few. While deformation has been studied for decades, deformable simulation is notorious for its costly computation. With the rapid development of sophisticated sensing devices and acquisition techniques, the complexities, scales and dimensionalities of the data have grown exponentially, and large-scale geometries are becoming ubiquitous in modern 3D data processing. Even with state-of-the-art hardware, a massive deformable simulation can still take hours, days, or even weeks. In this era of data explosion, increasing demands on both computing efficiency and simulation realism impose unprecedented challenges on this classic computing problem, so game-changing algorithmic techniques for large-scale, complex, and nonlinear deformable models are needed to empower future graphics applications. If successful, this project will not only expand the frontier of physics-based simulation technologies, but also profoundly inspire broader computing communities beyond graphics and enable a variety of applications. During a deformable simulation, a nonlinear system needs to be repetitively solved in order to track the continuous shape evolution of the deforming body. A deformable object with complex geometry could house a large number of unknown degrees of freedom, and the resulting high-dimensional integration becomes prohibitive. To overcome this problem, this project will develop a re-branded deformable model which systematically integrates advanced simulation techniques and deep learning (DL) tools, specifically deep neural networks (DNNs). The hypothesis is that digital simulation provides us nearly unlimited noise-free training data, which should be fully exploited and leveraged to benefit unseen yet difficult simulation or computing challenges. Unlike existing data-driven methods that interpret the data with a closed-form formulation (e.g., using a convex interpolation), DNNs provide a universal mechanism to extract intrinsic features hidden behind the raw data in an end-to-end manner, and have already demonstrated significant outcomes in many long-standing computer vision problems like object detection, classification, and annotation. However, harnessing DL in physics-based simulation is not easy. While in theory one may still encode all of these parameters using a very high-dimensional input vector, the corresponding network would be extremely large and complex. Even if we manage to collect sufficient training data to optimize this net, a single forward pass of it may be slower than a conventional simulator, making DL completely unprofitable. In this project, we will thoroughly investigate those grand technical challenges, forge a collection of data structures and algorithmic techniques for the data-driven deformable simulation, and thereby pave the way for DL-based physics simulation to next-generation computer graphics. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.