Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Learning 3D Structure from Video Streams
The goal of this project is to investigate how learning-based approaches can be employed to create a high quality, complete 3D reconstruction from a continuous image stream.
Keywords: 3D Reconstruction, Machine Leaning
3D Reconstruction from 2D images is one of the most fundamental problems in Computer Vision and Robotics, allowing robots to precisely navigate through the environment and to create high-quality digital representation of the environment. While over the past decades, approaches for 3D reconstruction were dominated by geometric methods, emerging learning-based techniques open new possibilities in this field [1], [2], [3]. Besides producing high-grade digital copies of a structure, learning-based approaches offer possibilities where geometric approaches reach their limits, such as reconstructing low-texture areas and occluded parts of the environment not directly seen by the camera. Therefore, this technology holds much potential to boost the quality of current 3D reconstruction techniques. The goal of this project is to investigate and develop such a learning-based pipeline for dense 3D reconstructing from a continuous image stream, that can produce a high-quality and complete reconstruction of the observed environment.
References:
[1] Xie et al., "Deep3D: Fully Automatic 2D-to-3D Video Conversion with Deep Convolutional Neural Networks", ECCV 2016
[2] Rezende et al., "Unsupervised Learning of 3D Structure from Images", NIPS 2016
[3] Han et al., "Image-based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era", T-PAMI 2019
[4] Golodetz et al., “Live Collaborative Large-Scale Dense 3D Reconstruction Using Consumer-Grade Hardware”, ISMAR-Adjunct 2018
3D Reconstruction from 2D images is one of the most fundamental problems in Computer Vision and Robotics, allowing robots to precisely navigate through the environment and to create high-quality digital representation of the environment. While over the past decades, approaches for 3D reconstruction were dominated by geometric methods, emerging learning-based techniques open new possibilities in this field [1], [2], [3]. Besides producing high-grade digital copies of a structure, learning-based approaches offer possibilities where geometric approaches reach their limits, such as reconstructing low-texture areas and occluded parts of the environment not directly seen by the camera. Therefore, this technology holds much potential to boost the quality of current 3D reconstruction techniques. The goal of this project is to investigate and develop such a learning-based pipeline for dense 3D reconstructing from a continuous image stream, that can produce a high-quality and complete reconstruction of the observed environment.
References:
[1] Xie et al., "Deep3D: Fully Automatic 2D-to-3D Video Conversion with Deep Convolutional Neural Networks", ECCV 2016
[2] Rezende et al., "Unsupervised Learning of 3D Structure from Images", NIPS 2016
[3] Han et al., "Image-based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era", T-PAMI 2019
[4] Golodetz et al., “Live Collaborative Large-Scale Dense 3D Reconstruction Using Consumer-Grade Hardware”, ISMAR-Adjunct 2018
Not specified
- Experience with machine learning tools
- Background in Computer Vision and 3D Reconstruction
- Excellent analytical skills, high self-motivation, research-driven attitude
- Programming experience (Python / C++)
- Experience with machine learning tools - Background in Computer Vision and 3D Reconstruction - Excellent analytical skills, high self-motivation, research-driven attitude - Programming experience (Python / C++)