What Happened in 3D


Loading...

Author / Producer

Date

2022

Publication Type

Master Thesis

ETH Bibliography

yes

Citations

Altmetric

Data

Abstract

Interpreting three-dimensional scenes that change over time is an open topic at its early stages. Current research mainly focuses on marking the existence of a difference between two temporal instances of the same scene without specifying the semantics of such difference. However, providing information on what has happened in an environment that changed over time can be helpful to a user only if the generated data is informative on the changes. In this project, I introduce Indoor Scene Activity Recognition, a new deep learning challenge that aligns elements of Action Classification and Change Detection. I annotate and collect a new dataset for this task by analyzing and manipulating another publicly available dataset, and I develop K3D, a dedicated two-stream three-dimensional neural network to tackle the challenge. Additionally, I create a two-dimensional convolutional baseline and an object multi-view baseline to benchmark K3D and compare the results in different settings.

Publication status

published

External links

Editor

Contributors

Examiner : Armeni, Iro
Examiner : Song, Shuran
Examiner : Pollefeys, Marc

Book title

Journal / series

Volume

Pages / Article No.

Publisher

ETH Zurich

Event

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Organisational unit

03766 - Pollefeys, Marc / Pollefeys, Marc check_circle

Notes

Funding

Related publications and datasets