Weak Supervision, n-shot Learning, Data Efficiency Broad motivation for this chapter: We want to use as little (annotated) data as possible, because data collection and data annotation are expensi...

Functional Analysis of Scenes Understanding possible interactions with objects (i.e. their function) Not the object type, but more: e.g. to know how to open a drawer, we also need to find its han...

Reconstructing and Generating Scenes Motivation: make shape generation more accessible, require less expertise capturing 3d photos (e.g. can change viewpoint a bit) 3d online in...

Object Detection and Instance Segmentation Idea now: Not only detect object classes, but learn to actually distinguish different objects (the chair on the left, the chair on the right, etc.). So h...

Semantic Scene Understanding - Semantic Segmentation 3D Semantic segmentation: segment scene (e.g. room) in semantic parts (e.g. for each part: to which furniture type it belongs) Popular Benchma...

Learning on Different 3D Representations Types of 3D Represenations Volumetric grids => 3D CNNs Multi-view Point clouds Meshes Volumetric Grids More efficient than dense grid? =>...

Shape Generation Goal: be able to generate shapes automatically. Usecases: for example, allow amateurs to create quality 3D models, and professionals to reduce repetitive work. Also, complete 3D s...

Shape Segmentation and Labeling 3D Classification tasks VoxNet [Maturana and Scherer ’15] NN architecture working on occupancy grids. One of the first 3D CNNs, used for object classification on oc...

Shapes - Alignment, Descriptors, Similarity 3D Shape Alignment (Registration) Often needed: e.g. in 3D scanning (combine different scanned parts) SLAM = Simultaneous Localization + Mapping (e...

Geometric Foundations - Surface Representations There are several ways to represent 3D shapes, each with their own advantages and disadvantages. Aspects to consider are the memory usage, how effic...