Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Deep Learning for Long-Term Place Recognition on UAVs
This goal of this project is to use deep learning techniques to learn data-driven place descriptors for long-term place recognition. We enforce one important criterion: the computation at runtime should be lightweight enough to run onboard a small UAV (Unmanned Aerial Vehicle).
Keywords: Place Recognition, UAV, Deep Learning, Machine Learning
This goal of this project is to use deep learning techniques to learn data-driven place descriptors for long-term place recognition. We enforce one important criterion: the computation at runtime should be lightweight enough to run onboard a small UAV (Unmanned Aerial Vehicle, such as the one de-picted above), without significantly sacrificing the accuracy or discriminative power of the de-scriptors.
Place recognition is the process of identifying whether a robot (e.g. a UAV) is revisiting a place that it has been to before. To answer this question, traditionally, we compare the current robot’s view to a database of images captured throughout the robot’s journey. However, as a particular place can undergo drastic perceptual changes because of variations in weather, seasons or illumination (e.g. day/night), images depicting the same location can be very different (as shown above) rendering long-term place recognition a very challenging problem. In this project, the aim is to study the image features that are used for navigation (e.g. binary feautres, such as ORB, BRISK, BRIEF) and beased on these, employ Deep Neural Networks (DNNs) to learn new features suitable for long-term place recognition; i.e. boosting their tolerance to lighting conditions.
Through this project, you will learn how a robot localization system works and how to design and implement your own deep learning architectures. Moreover, you will have the opportunity to employ your methodology on real setups and.equipment provided by the Vision for Robotics Lab. As this is a challenging project on the cutting edge of vision-based robotic perception, this project can lead to a submission to a top robotics conference or journal instead of a thesis report.
This goal of this project is to use deep learning techniques to learn data-driven place descriptors for long-term place recognition. We enforce one important criterion: the computation at runtime should be lightweight enough to run onboard a small UAV (Unmanned Aerial Vehicle, such as the one de-picted above), without significantly sacrificing the accuracy or discriminative power of the de-scriptors.
Place recognition is the process of identifying whether a robot (e.g. a UAV) is revisiting a place that it has been to before. To answer this question, traditionally, we compare the current robot’s view to a database of images captured throughout the robot’s journey. However, as a particular place can undergo drastic perceptual changes because of variations in weather, seasons or illumination (e.g. day/night), images depicting the same location can be very different (as shown above) rendering long-term place recognition a very challenging problem. In this project, the aim is to study the image features that are used for navigation (e.g. binary feautres, such as ORB, BRISK, BRIEF) and beased on these, employ Deep Neural Networks (DNNs) to learn new features suitable for long-term place recognition; i.e. boosting their tolerance to lighting conditions.
Through this project, you will learn how a robot localization system works and how to design and implement your own deep learning architectures. Moreover, you will have the opportunity to employ your methodology on real setups and.equipment provided by the Vision for Robotics Lab. As this is a challenging project on the cutting edge of vision-based robotic perception, this project can lead to a submission to a top robotics conference or journal instead of a thesis report.
The work to be undertaken involves five main work packages (WP):
• WP1: Literature review of existing convolutional neural network models, network pruning meth-ods and binary feature learning approaches
• WP2: Get started with some practical experience on existing code, and setting up the environ-ment to run real experiments
• WP3: Design and implementation of existing binary feature learning methods using deep learn-ing frameworks, such as Caffe, Tensorflow, Mxnet or Theano (the student can choose the framework they prefer).
• WP4: Train the model on our condition-varying dataset to generate condition-invariant binary Descriptors
• WP5: Conduct real experiments and comparisons to the state-of-the-art. Prepare the documen-tation and presentation of the project.
The work to be undertaken involves five main work packages (WP): • WP1: Literature review of existing convolutional neural network models, network pruning meth-ods and binary feature learning approaches • WP2: Get started with some practical experience on existing code, and setting up the environ-ment to run real experiments • WP3: Design and implementation of existing binary feature learning methods using deep learn-ing frameworks, such as Caffe, Tensorflow, Mxnet or Theano (the student can choose the framework they prefer). • WP4: Train the model on our condition-varying dataset to generate condition-invariant binary Descriptors • WP5: Conduct real experiments and comparisons to the state-of-the-art. Prepare the documen-tation and presentation of the project.
Desirable skills: -- cannot be required unfortunately in our department
Background basic knowledge in computer vision and machine learning
Programming experience in C++ and Matlab (or Python)
Experience in Linux is beneficial
Experience in one of the deep learning frameworks, such as Caffe, Mxnet or Tensorflow, is preferred, but not a requirement;
Desirable skills: -- cannot be required unfortunately in our department Background basic knowledge in computer vision and machine learning Programming experience in C++ and Matlab (or Python) Experience in Linux is beneficial Experience in one of the deep learning frameworks, such as Caffe, Mxnet or Tensorflow, is preferred, but not a requirement;
Zetao Chen, chenze@ethz.ch
Patrick Schmuck pschmuck@ethz.ch
Margarita Chli, chlim@ethz.ch
Zetao Chen, chenze@ethz.ch Patrick Schmuck pschmuck@ethz.ch Margarita Chli, chlim@ethz.ch