Structural and Compositional Learning on 3D Data (ICCV 2021 Workshop)

StruCo3D 2021


Computer Graphics Computer Vision & Pattern Recognition



3D structure and compositionality lie at the core of many methods for different tasks in computer vision, graphics and robotics, including but not limited to recognition, reconstruction, generation, planning, manipulation, mapping and embodied perception. Unlike traditional connectionist approaches in deep learning, structural and compositional learning includes components that lean more towards the symbolic end of the spectrum, where data or functions are represented by a sparse set of separate and more clearly defined concepts. For example, in 3D objects, this could be a decomposition of an object into spatially localized parts and a sparse set of relationships between them, or in scenes, it could be a scene graph, where rich inter-object relationships are described. Similarly, a navigation or interaction task in robotics can also be decomposed into separate parts of concepts or submodules that are related by spatial, causal, or semantic relationships.
People from different fields or backgrounds use different structural and compositional representations of their 3D data for different applications. We bring them together in this workshop to have an explicit discussion of the advantages and disadvantages of different representations and approaches, as well as to share, discuss and debate the diverse opinions regarding the following questions:
- Which types of structure should we use for different tasks and applications in graphics, vision and robotics?
- How should we factorize a given problem into sparse concepts that make up the structure?
- How should we factorize different types of 3D data into sparse sets of components, relationships, or operators?
- Which algorithms are best suited for a given type of structure?
- How should we mix structural and non-structural approaches?
- Which parts of a problem are suited for structural approaches, and which ones are better handled without structure?
We accept both archival full paper (up to 8 pages) and non-archival short paper (up to 4 pages) submissions. Every accepted paper will have the opportunity to give a 10-min spotlight presentation and host two 30-min poster sessions (12-hours separated).
Please refer to the workshop website and CFP page for more details: https://geometry.stanford.edu/struco3d
Contact: struco3d@googlegroups.com or kaichun@cs.stanford.edu