Schedule
Opening remark
Workshop organizers
Time: 8:30-8:40am
Implicit Neural Representations
Yaron Lipman (Weizmann Institute of Science)
Time: 8:40 am - 9:30 am, June 19th
Talk slides: PDF
Abstract: In this talk we will discuss novel methods for modeling geometry implicitly using neural networks. Namely, representing 3D surfaces as zero level-sets of neural networks. We will discuss different approaches to control and optimize such representations as well as cover related applications including: surface reconstruction from raw 3D data, shape space learning, and differentiable rendering.
Learning 3D Reconstruction in Function Space
Andreas Geiger (University of Tuebingen and Max Planck Institute for Intelligent Systems)
Time: 9:30 am - 10:20 am, June 19th
Talk slides: PDF
Abstract: In this talk, I will show several results of my research group on learning neural implicit 3D representations by departing from the traditional concept of representing 3D shapes explicitly using voxels, points and meshes. Implicit representations have a small memory footprint and allow for modeling arbitrary 3D toplogies at (theoretically) arbitrary resolution in continuous function space. I will show the ability and limitations of these approaches in the context of reconstructing 3D geometry (CVPR 2019) and textured 3D models and motion (ICCV 2019). Finally, I will also show very recent results (CVPR 2020) on learning implicit 3D models using only 2D supervision by deriving an analytic closed form solution to the gradient updates.
Learned Embeddings for Geometric Data
Isaak Lim (RWTH Aachen University)
Time: 10:20 am - 11:10 am, June 19th
Talk slides: PDF
Abstract: Ideally, we would like to plug two geometric objects into a function, which then tells us the similarity between them. This would allow us to answer a variety of questions on several different levels about geometric data in down-stream applications. However, for high-level tasks such as computing style similarity or vertex-to-vertex maps between 3D shapes it is difficult to do so directly on raw geometric data, since more structured aggregated information is needed for more abstract tasks. One way of implementing such a similarity function is to first compute maps of this data to an embedding space, which en-codes meaningful relations between different geometric elements, e.g. stylistically more similar shapes are embedded closer together. By making use of this embedding space we are able to calculate and output a similarity measure. However, it is difficult to construct maps that preserve such properties by hand because it becomes increasingly challenging to formulate explicit rules or models for evermore abstract tasks. For this reason we make use of geometric data collections with task-related meta information provided by humans. This allows us to formulate the map computation in flexible terms by using neural networks without making too many assumptions about the form of the map itself. In order to benefit from a wide range of available machine learning techniques, we have to first consider the problem on how to choose a suitable representation of geometric data as input to various learning models. Concretely, depending on the availability of the data sources and the task specific requirements we compute embeddings from images, point clouds and triangle meshes. Once we have found a suitable way to encode the input, we explore different ways in which to shape the learned intermediate domain (embedding) that ex-tend beyond straightforward cross-entropy minimization based approaches on categorical distributions. We show that these approaches lend themselves to both discriminative as well as generative tasks.
Hybrid Representations for 3D Understanding
Qixing Huang (University of Texas at Austin)
Time: 11:10 am - 12:00 pm, June 19th
Talk slides: PDF
Abstract: Choosing suitable data representations is one of the most critical topics when installing machine learning on 3D data. Instead of exploring a single representation, there is great potential in integrating multiple representations. Such hybrid representations exhibit advantages in uncertainty reduction, adaptative feature selection, and self-supervision, among others. This talk introduces several recent results in pose estimation, object detection, and 3D reconstruction.
Recognizing Objects from Any View with Object and Viewer-Centered Representations
Sainan Liu (University of California, San Diego)
Time: 12:00 pm - 12:25 pm, June 19th
Talk slides: PDF
Abstract:
Objects are three-dimensional in the physical world, but the recognition tasks in computer vision have been primarily performed on 2D natural images. When training a network with a limited number of views of an object instance, it may have a hard time recognizing the same object instance from an unseen viewpoint. In this talk, I will talk about our recent work that improves object recognition from any view using both object and viewer-centered representations.
12:25pm -- 1:30pm Break
Geometric deep learning for 3D human body synthesis
Michael Bronstein (Imperial College London)
Time: 1:30 pm - 2:20 pm, June 19th
Abstract: Geometric deep learning, a new class of ML methods trying to extend the basic building blocks of deep neural architectures to geometric data (point clouds, graphs, and meshes), has recently excelled in many challenging analysis tasks in computer vision and graphics such as deformable 3D shape correspondence. In this talk, I will present our recent research efforts in 3D shape synthesis, focusing in particular on the human body, face, and hands.
More from Less: Reducing Supervision for 3D Shape Segmentation
Siddhartha Chaudhuri ( Adobe Research & IIT Bombay)
Time: 2:20 pm - 3:00 pm, June 19th
Talk slides: PDF
Abstract: Shape segmentation is a foundational problem in geometry processing, with a rich history of techniques using both traditional analytical optimization and modern learning-based methods. The advantage of approaches using statistical learning is that they can leverage complex priors that are better inferred from exemplars than written by hand. The disadvantage is that such priors typically require extensive supervision, i.e. very large collections of segmented and labeled exemplars, to train effectively. Such annotation is tedious and expensive. This talk will present a body of work, that we did over the last 3 years, to progressively reduce the amount of supervision required for 3D single-shape segmentation. I will begin with a brief review of our initial work on fully-supervised segmentation [2017], and then move on to weakly-supervised, one-shot, and unsupervised methods [2018-2020]. I will talk about state-of-the-art (at the time of publication) techniques in all of these cases, and conclude with a summary of pros and cons, lessons learned, and directions for future work.
Learning Shape Functionality and Part Mobility
Ruizhen Hu ( Shenzhen Univ.)
Time: 3:00 pm - 3:45 pm, June 19th
Talk slides: PDF
Abstract: One of the goals of computer graphics is to provide tools for designing and simulating real or imagined artifacts, such as man-made objects. Such artifacts are usually functional, and thus an understanding of functionality is paramount for simulating and validating different designs of artifacts. In recent years, computer graphics and related fields, such as computer vision and robotics, have also devoted much attention to the inference of possible motions of 3D objects and their parts, since this problem is closely related to an understanding of object and functionality. In this talk, I will introduce recent works on analysis and understanding shape functionality and part mobility using deep learning-based methods.