Robust pose estimation of objects is needed for robotic application and human computer interfaces. Like most computer vision tasks, robust pose estimation becomes a challenging problem when the environmental conditions (like lighting and visibility) are not controlled.
The approach we propose is to leverage the computational graphs that can be created in deep learning framework like TensorFlow. This approach opens up several strategies.
For example, if a model of an object is available, a deep neural network can be trained to make a first guess for the pose of the object. This first guess is then refined by optimizing the re-projection error of the current guess by comparing the edge image of the actual view and the virtual view of the guess.
As part of the research project, you will be involved in:
- creating a data set of artificial images and real images
- creating and training neural networks for the first guess prediction
- designing and implementing optimisation algorithms for the pose refinement phase
- testing and validating the systems that are developed.
As a result of this project, we expect to develop a prototype of a vision system for robust pose estimation
Skills and experience
To be considered for this project, it is essential that you have excellent programming and mathematics skills. It is also desirable if you have prior knowledge of Python and machine learning.
Contact the supervisor for more information.