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-Learned Constellation-based descriptors for Semantic SLAM
Using deep learning to generate constellation-based descriptors for a more robust representation of the environment in semantic SLAM
Keywords: Deep Learning, Semantics, SLAM, Descriptors
Robots become more and more suited for everyday tasks around humans. This requires the robot to be able to navigate in a non-controlled, highly dynamic environment. This becomes especially challenging outdoors with vast changes in appearance due to daytime, weather and season. Thanks to the recent advantages of deep learning for object recognition tasks also in those challenging environments, semantic objects become more and more viable to represent the environment. Unfortunately, often those objects do not offer the required amount of descriptiveness to reliably re-discover them in a scene. However, their constellation within a local neighborhood of other objects is often more distinct.
In this project, we aim to investigate how semantic objects can be described with their constellation among other nearby objects. This includes the development of a data generation pipeline in order to train and test a new neural network structure for constellation based descriptors. Furthermore, our chosen approach should be compared to a hand-crafted descriptor baseline. Ultimately, the approach is evaluated both in simulation and with real-world data by including it in a state-of-the-art mapping framework.
Robots become more and more suited for everyday tasks around humans. This requires the robot to be able to navigate in a non-controlled, highly dynamic environment. This becomes especially challenging outdoors with vast changes in appearance due to daytime, weather and season. Thanks to the recent advantages of deep learning for object recognition tasks also in those challenging environments, semantic objects become more and more viable to represent the environment. Unfortunately, often those objects do not offer the required amount of descriptiveness to reliably re-discover them in a scene. However, their constellation within a local neighborhood of other objects is often more distinct.
In this project, we aim to investigate how semantic objects can be described with their constellation among other nearby objects. This includes the development of a data generation pipeline in order to train and test a new neural network structure for constellation based descriptors. Furthermore, our chosen approach should be compared to a hand-crafted descriptor baseline. Ultimately, the approach is evaluated both in simulation and with real-world data by including it in a state-of-the-art mapping framework.
- Literature review for semantic SLAM and descriptors for objects.
- Development of a data generation pipeline for different object constellations.
- Design, implementation and training of a deep neural network for constellation descriptors.
- Evaluation of descriptor performance on both simulated and real-world data.
- Integration in state-of-the-art mapping framework.
- Literature review for semantic SLAM and descriptors for objects. - Development of a data generation pipeline for different object constellations. - Design, implementation and training of a deep neural network for constellation descriptors. - Evaluation of descriptor performance on both simulated and real-world data. - Integration in state-of-the-art mapping framework.
- Highly motivated and independent student.
- Interest in computer vision and deep learning.
- Excellent programming skills (Python) mandatory.
- Experience with object detection or learning preferable.
- Experience with SLAM and ROS preferable
- Enrolled at ETH Zurich.
- Highly motivated and independent student. - Interest in computer vision and deep learning. - Excellent programming skills (Python) mandatory. - Experience with object detection or learning preferable. - Experience with SLAM and ROS preferable - Enrolled at ETH Zurich.
Please send your cv and transcripts to: Andrei Cramariuc (andrei.cramariuc@mavt.ethz.ch) and Florian Tschopp (florian.tschopp@mavt.ethz.ch)
Please send your cv and transcripts to: Andrei Cramariuc (andrei.cramariuc@mavt.ethz.ch) and Florian Tschopp (florian.tschopp@mavt.ethz.ch)