You can find the detailed descriptions of the five open positions in the 3D Media Group and CIVIT below. The BSc and MSc Students will be supervised by the 3D Media Group researchers. For more details, please contact the contact persons mentioned under each position. Submit your application here, and don’t forget to select the desired position(s) there. After submitting your application, you might want to drop an email to the contact persons for each position you have applied for, to inform them about this. This will help us to track your application easier.
Computational hyperspectral imaging
Position number 28
Short description: Conventional color cameras integrate the visible spectrum within three different bands (color channels), using red, green, and blue color filters. Hyperspectral (HS) cameras can record much denser bands, which in turn provide richer information regarding the material properties and the lighting. Such information is beneficial in various areas such as remote sensing, medical imaging, and computer vision. However, capturing HS data with single shot is challenging due to its complexity. The compressive HS cameras address this challenge with the help of coded measurements and post-processing.
We aim at designing novel computational HS cameras utilizing state-of-the-art machine learning (ML) algorithms. The general approach is illustrated in the above figure. Your tasks within the project include:
Stage 1
- Brief overview of HS image representations (aka HS datacube) and acquisition.
- Implementation of machine learning algorithms for optimizing (training) our existing computational HS camera model.
Based on the progress and results of the tasks in Stage 1, we would like to proceed further with the following tasks, which are expected to lead to a MSc. thesis:
Stage 2
- Comprehensive review and implementation of the current state-of-the-art spectral image reconstruction algorithms based on machine/deep learning.
- Getting familiar with rigorous simulation of imaging, including modelling and optimization of optics and sensor components. Design and implementation of a computational HS imaging system utilizing machine/deep learning.
- Collecting HS data in various imaging conditions using the spectral-scanning cameras available in Centre for Immersive Visual Technologies (CIVIT). Such data is planned to be utilized as ground truth data for machine learning based HS imaging algorithms; it is expected to be beneficial to the research community world-wide, as currently only limited ground truth HS data is publicly available.
Fulltime position: YES
Start Date: Beginning of summer, or upon agreement. Earlier is also possible.
End Date: The project (Stage 1) is intended for 3 months. An extension (Stage 2) is possible to result in MSc. thesis work.
Prerequisites: Programming skills (MATLAB/Python), basics of image processing, basic knowledge of machine learning and CNNs. Knowledge of optics is a plus. Trainee position targeted for MSc. Students.
Research environment: The research assistant is to be integrated within 3D Media Group (https://research.tuni.fi/3dmedia/) and use the equipment of the Centre of Immersive Visual Technologies (www.civit.fi).
Contact person email: Ugur Akpinar: ugur.akpinar@tuni.fi, Erdem Sahin: erdem.sahin@tuni.fi
Machine learning for (coherent) computational near-eye displays
Position number 29
Short description: The end-to-end optimization of a coherent imaging system involves learning the optimal optical and data processing setup from large image data sets for a given task. One example of such a learning task is the optimization of the optics and pre-processing of data simultaneously for a hypothetical coherently illuminated near-eye display (see the figure below). In a conventional near-eye display setup, the optics (i.e. a lens) positions the displayed content at a further depth from the viewer. Assuming a simple lens, the image is sharp only at a single depth and blurred at all other depths according to the basic principles of lens imaging. By replacing the optics and applying a pre-processing step with the results of an end-to-end optimization framework, the combined effect produces images that are sharp either at any arbitrary depth (accommodation-invariant (AI)), at specified depths (multifocal) or dynamically blurred based on the displayed content (accommodation-enabled).
This project utilizes an end-to-end optimization framework for AI and multifocal near eye displays being developed at 3D Media Group. Currently, the framework utilizes image data only as its input, and thus, the imaging model assumes the same depth across the entire image. The goal is to enhance the framework by introducing depth information (i.e. depth maps) in addition to the images as its input, thus requiring appropriate modelling for the imaging process. This can improve the existing optimization results of AI and multifocal displays. Furthermore, augmenting the framework in this way enables the efficient use of dynamic optics to obtain the accommodation-enabled display design.
Fulltime position: YES
Start date: Beginning of summer. Upon agreement, an earlier start is also possible.
End date: The project is intended for 3 months; extension is possible to result in MSc. thesis work.
Prerequisites: MATLAB programming skills, basics of image processing, basic knowledge of machine learning and CNNs. Knowledge of optics is a plus. Trainee position targeted for MSc. students.
Research environment: The research assistant is to be integrated within 3D Media Group (https://research.tuni.fi/3dmedia/) and to use the equipment of the Centre of Immersive Visual Technologies (www.civit.fi).
Contact person email: Jani Mäkinen: jani.makinen@tuni.fi, Erdem Sahin: erdem.sahin@tuni.fi
3D volumetric reconstruction in Light Field Microscopy
Position number 30
Short description: Classical microscopy requires hundreds of measurements (images) for the 3D reconstruction of a specimen captured by a microscope. In contrast, light field (LF) microscopy can reconstruct the specimen from a single capture by means of computational imaging. This is possible due to the properties of the LF, which is multidimensional image data which contains both spatial and angular information. This facilitates the reconstruction of the distance to the individual points in the visual scene. Currently available reconstruction algorithms rely on solving the underlying ill-posed inverse reconstruction problem. However, due to many issues such as object transparency, small spatial details, etc. these solutions are not optimal and there is a need for improvements.
The goal of this project is to develop machine learning methods for volumetric reconstruction from LF microscopy data. The following tasks are planned:
- Literature overview about principles of LF microscopy and state-of-the art reconstruction methods.
- Generation of appropriate synthetic LF datasets simulating capture by various types of LF microscopes.
- Designing and training convolutional neural network models for volumetric reconstruction of specimens from LF images
- Comparison of achieved results against state-of-the-art algorithms.
The scope of the planned work, assuming all tasks are successfully completed, is enough for a MSc thesis.
Fulltime position: Full time summer internship
Start date: May-June (start date is flexible)
End date: 3-4 months during summer with possible continuation on part-time basis during next academic year
Prerequisites: Good programming skills in Matlab & Python. Knowledge of signal/image processing and machine learning is required.
Research environment: 3D Media Group (https://research.tuni.fi/3dmedia/), Centre for Immersive Visual Technologies (www.civit.fi)
Contact person email: Robert Bregovic: robert.bregovic@tuni.fi
Machine learning for light field displays
Position number 31
Short description: A light field (LF) display is a 3D display which aims at recreating the complete LF corresponding to a 3D scene. This results in a realistic 3D visualization and providing immersive visual experience. Due to the current technological limitations, existing LF displays can generate only a limited number of light rays, which for a small subset of the complete LF. In order to improve the visual appeal of the scene visualized by the display, the values assigned to those light rays must be properly optimized based on the display model and the content to be visualized. Existing optimization methods are typically tailored to a particular type of LF display (e.g. Projector-based LF displays, Integral imaging LF displays). These methods fail short in providing a generalized formulation that would be applicable for any type of LF display.
The goal of this project is to derive and implement a machine learning framework for optimizing the input images of various LF displays. The following tasks are planned for achieving this goal:
- Study the in-house developed optimization framework for Integral-imaging LF displays based on machine learning.
- Take the in-house developed ray-based model for projector-based LF display and apply machine learning for optimizing the LF display output.
- Implement a common framework for optimizing the LF display output for projector-based and integral-imaging LF displays.
Fulltime position: Full time summer internship
Start date: March-June (preferably as early as possible)
End date: 3-4 months during summer with possible continuation on part-time basis during next academic year
Prerequisites: Mandatory: machine learning, Python, C++, Matlab. Recommended: statistical data analysis (regression, scatterplots, etc.), 3d rendering, 3d image processing.
Research environment: 3D Media Group (https://research.tuni.fi/3dmedia/); Centre for Immersive Visual Technologies (www.civit.fi)
Contact person email: Robert Bregovic: robert.bregovic@tuni.fi
Measuring and collecting room impulse responses using robotic arm
Position number 32
Short description: This project involves collecting impulse responses from different spaces in the university. Involves the control and usage of robotic arm.
Suitable for MSc student with automation background. Prior knowledge in robotic manipulation is required. Details involving impulse response measurements will be provided by us.
Fulltime position: YES
Start date: June-2021
End date: August -2021
Prerequisites: Knowledge in MATLAB scripting and execution.
Contact persons emails: Arjun Venkatakrishnan: arjunvenkat.venkatakrishnan@tuni.fi and Pasi Pertilä: pasi.pertila@tuni.fi