The master's thesis projects are a very important part of exploratory work at Imint. Every year we invite a number of students to perform master's thesis projects in areas we find to be of interest for future opportunities. Many of our previous thesis students have also started working with us after finishing their master's thesis.
Here you will find the thesis topics we have available for 2022, and you're welcome to apply using the "apply for this job" button. Please specify which topic or topics you are interested in when writing your application.
Real-time fully automatic calibration of camera parameters
When performing certain image and video processing, it is often important to know some of the intrinsic camera parameters, such as focal length, lens distortion or exposure time. Additionally, motion sensors, such as a gyroscope or an accelerometer, are often used together with the camera and these have their own set of parameters. The parameters are usually known to some degree, but small variations between devices and how it is reported means it often needs to be calibrated anyway, which is a time consuming process. Imint has extensive calibration experience and a full set of tools for this purpose, but it is still a manual process. The goal of this thesis is to develop an algorithm that lets a phone calibrate itself, with the available tool set, when a user is capturing normal video. The code will be written in C++ in Imint's video processing framework.
Video sub-pixel frame alignment
When a camera moves while capturing a video, there will be a lot of random movement between each frame capture. An important aspect of video stabilization is to map each frame onto the previous, after which a more steady camera path can be created. This is usually accurate down to the level of a few pixels, which is enough for video stabilization. There are other video processing products, however, that need higher accuracy. The goal of this thesis is to develop a sub-pixel accurate frame to frame mapping, using an existing relatively accurate starting point, that can run in real-time on a smartphone. The code will be written in C++ in Imint's video processing framework.
Real-time video HDR
HDR (high dynamic range) imaging is a concept where we increase the range of brightness that a camera can capture beyond what the physical sensor is capable of. This is especially useful for smaller sensors, such as the ones found in smartphones. It can be done in several ways. The classic way is to capture several pictures at different exposures and combine them. For video though, this method is difficult to use due to the limited time available between frames. The objective of this thesis is to develop an HDR algorithm that works for video and can run in real-time on a smartphone. The code will be written in C++ in Imint's video processing framework.
Real-time video super resolution
Super resolution is an important technology for smartphones where the optical zoom options are limited. For these phones, digital zoom is applied when the user zooms in, which tends to look very poor with classic interpolation techniques. Super resolution has come a long way and is now commonly available for photos on flagship smartphones. We have a foundation for super resolution at Imint, and we are now looking beyond photos. The objective of this thesis is to adapt existing super resolution research and algorithms to make it run in real-time for video. The code will be written in C++ in Imint's video processing framework.
Local motion compensation
A normal user made video has a mix of random and intended movement, both from the camera and movement in the scene. The foundation for video stabilization is to compensate for the camera movement by warping each frame to match with the previous frames. We are now looking at the movement within the scene, and how it can be calculated. The objective of this thesis is to develop an algorithm that can warp a frame to compensate for people and objects moving during a video. The code will be written in C++ in Imint's video processing framework.
Real-time video spatial noise reduction
Spatial noise reduction is a huge research field, with a breadth of different solutions. We are now looking for the best way to perform spatial noise reduction of real-time video on smartphones. The objective of this thesis is to do a research review of state of the art methods, creating a proof of concept of the most interesting options, as well as modifying the algorithm to fit real-time implementation on a smartphone. The code will be written in C++ in Imint's video processing framework.
Real-time video bokeh
Artificial bokeh for photos is more or less a standard feature on flagship smartphones, usually implemented in a portrait mode. We are now looking into implementing a bokeh function for video, which will significantly increase the requirements on processing time. We assume that a low resolution depth map and face detection output are already available. The objective of this thesis is two-fold. The first part is to figure out a way to do a high resolution segmentation of a person, and the second part is to filter only the background in a fast but realistic looking way. We could also imagine the thesis focusing on only one of these parts, or having two students cooperating by focusing on each part individually. The code will be written in C++ in Imint's video processing framework.
If you have questions regarding any of these projects, we're happy to answer them. Please connect with us on our career portal, career.weareimint.com, to start a conversation.