CESS seminar - Generating Semantic Building Information Over Time and Space Using Visual Data
Abstract
Many construction and operation processes, such as those related to sustainable construction and circular economy, require information on the state of the building (spatial), as well as on how it became what it is today (temporal). State-of-the-art approaches attempt to acquire spatial and temporal building information using easily obtainable visual data and Computer Vision algorithms. However most of them examine a static version of our world when for many real-world applications we require a spatiotemporal interpretation of it. The examined change, especially in the 3D domain, is usually constrained to small spatial (e.g., that of a room) and temporal (e.g., that of a few minutes or seconds) scales, and mainly relates to relocation (movement of humans and objects). Considering the lifecycle stages of the built environment, from construction to operation and end-of-use, consisting elements undergo changes that extend beyond relocation and relate to differences in the elements’ geometry, appearance (e.g., floors before and after carpet is added), and topology (e.g., walls begin as a group of studs and gradually change into the final wall structure that users see). To address this, I am employing the challenging setup of construction sites (new or under renovation) as the grounds for developing computer vision methods that are able to work even in the most drastic of changes. The first step is to register 3D point clouds of the same space captured over time since they are almost never aligned in the same coordinate system. In this talk I will present a benchmark and dataset for automatic partial and multi-way registration of the data regardless at which point in time they were captured. This is a very challenging scenario for computer vision algorithms due to the drastic changes in geometry, appearance, and topology that take place over time, as well as the repetitive structure of these scenes. I demonstrate the need for new algorithms that can handle such challenging scenes and drastic changes, as well as the relevance of the dataset.
Short biography
Iro Armeni is a PostDoctoral Researcher at ETH Zurich and an incoming Assistant Professor at Stanford University (Sept. ’23), conducting interdisciplinary research between Architecture, Civil Engineering, and Visual Machine Perception. Her area of focus is on automated semantic and operational understanding of buildings throughout their life cycle using visual data. She completed her PhD at Stanford University on August 2020, Civil and Environmental Engineering Department, with a PhD minor at the Computer Science Department. Prior to enrolling in the PhD program, Iro received an MSc in Computer Science (Ionian University-2013), an MEng in Architecture and Digital Design (University of Tokyo-2011), and a Diploma in Architectural Engineering (National Technical University of Athens-2009). She is the recipient of the ETHZ Postdoctoral Fellowship, the Google PhD Fellowship on Machine Perception, and the Japanese Government (MEXT) scholarship. Iro has worked as an architect and consultant for both the private and public sector.
Many construction and operation processes, such as those related to sustainable construction and circular economy, require information on the state of the building (spatial), as well as on how it became what it is today (temporal). State-of-the-art approaches attempt to acquire spatial and temporal building information using easily obtainable visual data and Computer Vision algorithms. However most of them examine a static version of our world when for many real-world applications we require a spatiotemporal interpretation of it. The examined change, especially in the 3D domain, is usually constrained to small spatial (e.g., that of a room) and temporal (e.g., that of a few minutes or seconds) scales, and mainly relates to relocation (movement of humans and objects). Considering the lifecycle stages of the built environment, from construction to operation and end-of-use, consisting elements undergo changes that extend beyond relocation and relate to differences in the elements’ geometry, appearance (e.g., floors before and after carpet is added), and topology (e.g., walls begin as a group of studs and gradually change into the final wall structure that users see). To address this, I am employing the challenging setup of construction sites (new or under renovation) as the grounds for developing computer vision methods that are able to work even in the most drastic of changes. The first step is to register 3D point clouds of the same space captured over time since they are almost never aligned in the same coordinate system. In this talk I will present a benchmark and dataset for automatic partial and multi-way registration of the data regardless at which point in time they were captured. This is a very challenging scenario for computer vision algorithms due to the drastic changes in geometry, appearance, and topology that take place over time, as well as the repetitive structure of these scenes. I demonstrate the need for new algorithms that can handle such challenging scenes and drastic changes, as well as the relevance of the dataset.
Short biography
Iro Armeni is a PostDoctoral Researcher at ETH Zurich and an incoming Assistant Professor at Stanford University (Sept. ’23), conducting interdisciplinary research between Architecture, Civil Engineering, and Visual Machine Perception. Her area of focus is on automated semantic and operational understanding of buildings throughout their life cycle using visual data. She completed her PhD at Stanford University on August 2020, Civil and Environmental Engineering Department, with a PhD minor at the Computer Science Department. Prior to enrolling in the PhD program, Iro received an MSc in Computer Science (Ionian University-2013), an MEng in Architecture and Digital Design (University of Tokyo-2011), and a Diploma in Architectural Engineering (National Technical University of Athens-2009). She is the recipient of the ETHZ Postdoctoral Fellowship, the Google PhD Fellowship on Machine Perception, and the Japanese Government (MEXT) scholarship. Iro has worked as an architect and consultant for both the private and public sector.
Practical information
- Informed public
- Free
Organizer
- Prof. Olga Fink (IMOS), Prof. Alex Elahi (VITA), Prof. Dusan Licina (HOBEL) and Prof. Alain Nussbaumer (RESSLab)
Contact
- Prof. Stefana Parascho (CRCL)