MSc and BSc Thesis Topics Available

The following topics are available to TUT students interested to join the Vision Group. All these topics can be:
  • A course project - a miniature proof of concept or at least a proper try wink
  • BSc thesis - a miniature proof of concept with examples
  • MSc thesis - a proof of concept with examples and small quantitative experiments
  • PhD? thesis - a proof of concept with amazing demonstrations and extensive quantitative experiments
Please, note that these are not necessarily paid jobs, but provide interesting topic to finish your studies and demonstrate your skills!

Specific available topics are the following:

Cascading Convolutional Neural Networks

Our laboratory has recently developed a novel algorithm for building strong cascades. During this project you will implement the algorithm for cascading the state-of-the-art CNNs (e.g., AlexNet? and VGGNet). You will study whether our cascading method can improve the results in ImageNet? classification.

deepnn_yuan.png
  • Strong programming skills (Matlab/C/C++/Python) in Linux are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Biomedical imaging: mammographic image analysis

The aim of this project is the development of new tools for the analysis of mammography images for breast cancer detection and diagnosis. Specifically, we are currently developing an automatic system for mammography image analysis based on the breast anatomy. For this work, you will learn different techniques and tools for image analysis and biomedical research.


stmapping.png
  • Matlab skills required
  • Supervisor: Said Pertuz (said.pertuz@tut.fi)
  • Contact supervisor for details

Lightfield processing

Lightfield cameras (also known as plenoptic cameras) are a new technology that is expected to revolutionize digital photography. Plenoptic cameras allow for the generation of post-capture effects such as digital refocusing, 3D stereo rendering and 3D scene recontruction, among others. It is new and exciting emerging field in computer vision and image processing. For this project, you will learn classical image processing techniques such as depth estimation and segmentation and will apply them to these new devices.

Lytro_ILLUM_DSC_7403-.jpg
  • Matlab skills required
  • Supervisor: Said Pertuz (said.pertuz@tut.fi)
  • Contact supervisor for details

Build Own 360-Degree Video Camera

In this work, you will compete with commercial 360-degree video cameras such as Richo Theta, Samsung Gear, or Nokia OZO and build your own from cheap Web cams. You will learn about 3D computer vision for camera calibration and about fast mappings to generate the 360-degree view point on-the-fly.

360_image_small.jpg

  • Strong programming skills (Matlab/C/C++/Python) in Linux are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Light & Magic: Representing Motion in a Still Image

In cartoons motion is represented by lines that show, for example, how a hand moves (e.g., Donald Duck hits a tennis ball), but your challenge in this topic is how to create a visually plausible motion experience from a video to a single image. You will study the literature and finally implement code that takes a short video as an input and generates a still image the encodes motion of the video. This topic is for students who wish to combine art and computer science!

hs_joni_magic_wand_small.jpg
  • Strong programming skills (Matlab/C/C++/Python) in Linux are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Differential Evolution Training Algorithm for Deep Convolutional Neural Networks

In this project you will study two hot topics of artificial intelligence: deep convolutional neural networks (DCNNs) and genetic algorithms (GAs). In particular, you will extend our highly cited differential evolution training algorithm (see pdf ) to DCNNs and compare it to the stochastic gradient descent for the ImageNet? dataset. During this thesis you will become a master of knowledge on deep neural networks and if successful headhunted by Google, Facebook, Amazon, Baidu, Microsoft etc. wink

deepnn_yuan.png
  • Strong programming skills (Matlab/C/C++/Python) in Linux are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Deformable Part Model for ImageNet? Large Scale Visual Recognition Challenge Detection

You will familiarize yourself with one of the most well known methods for visual class detection, the Deformable Part Model (DPM) by Pedro Felzenszwalb and his group (see DPM page for references, in particular, the Rob Girshick's PhD? thesis). Your task is first to replicate the Pascal VOC learning and detection as described in the DPM V5.0 Github and then write scripts to run the code for 200 classes in the ImageNet LSVRC 2014 (i.e., files imagenet.m, imagenet_train.m etc.) Moreover, you will write suitable scripts to run code parallel in the TUT distribute computing system Merope . As the last important point you will investigate the effects of the DPM parameters for the detection accuracy.

dpm_detection.jpg
  • Strong programming skills (Matlab) are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Image Captioning for Optical Character Recognition

Traditional optical character recognition (OCR) is based on segmenting the symbols and passing each to a classifier. There are two key steps, where this process may fail: segmentation and parsing of the results. Both these problems can be avoided using the recently introduced Image Captioning approach. The standard use of image captioning transforms images into textual descriptions (and obviously requires huge amounts of training data). In this project, we study a simpler and less ambiguous problem: how to map a picture of text into a text string using the recently introduced combination of a deep convolutional neural network with a recurrent LSTM network.
  • Skills: Python or strong interest to learn it. We use the Keras platform for implementation.
  • Supervisor: University Lecturer Heikki Huttunen
  • Contact supervisor for more details.

Big Data with Support Vector Machines in Distributed Computing Systems

In this project you will need to make a novel parallel and incremental training system that allows using the powerful support vector machine (SVM) classifiers for big data problems - huge amount of data and huge amount of classes. The main tools will be the libSVM package and the slurm resource allocation system in Linux. This project is for those who wish true coding project to test their skills. If you're not a proper nerd, don't even think to try this!

big_data_flickr_infocux_small.jpg
  • Hacker level programming skills (C/C++/Python) and Linux knowledge are must!
  • Super hacker: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Light & Magic: Harry Potter's Wand

You'll need to implement an interactive computer vision system that follows a stick in the users hand and uses a video projector to create visual effects. For example, the "wand-lighting charm" would make the projector to produce a tiny white spot that follows the light. Moreover, "Avada Kedavra" (killing curse) would create green lights from the wand.

potters_wand.jpg
  • Strong programming skills (C/C++/Python) in Linux are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

(Tiny) ImageNet? Classification Using Deep Neural Networks

This is a very trendy topic and currently under active research in the top computer vision labs in the top universities and also in the research centres of the top companies such as Google, Facebook and Microsoft. The challenge is to classify a huge amount of images to 1,000 pre-defined classes (see http://image-net.org/challenges/LSVRC/2013/). The state-of-the-art technique is via deep neural networks learning (see http://caffe.berkeleyvision.org /). In this project, you play with a "tiny" version of the ImageNet? (see https://bitbucket.org/kamarain/imagenet-tiny) which is much faster to train and use (you have full control to the size of images). You will also play with different deep neural network architectures available in Caffe. Starting code will be provided, but the work consists of a lot of playing with networks and their parameters and your challenge is to beat the state-of-the-art results in the ILSVRC 2013/2014 Web site.
  • Strong programming skills are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Pattern Recognition Using Random Gaussian Mixture Models

Randomisation is a trendy trick used in machine learning during the recent years. Typically it has been used with discriminative methods and special approaches such as to extend the idea of decision tree learning to many "kind of randomly" trained trees leading to the concept of random forests. Our case is a bit different, Gaussian mixture models (GMMs) are powerful tools to model probability densities and we have developed a popular Matlab toolbox GMMBayes to train and use GMMs. Your task is to "randomize" Gaussian mixture models to make them to cope with a small amount of data, for example.
  • Matlab skills are required.
  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

Automatic 3D Scene Reconstruction From a Single Image

In his MSc thesis Teemu Tarkiainen developed a tool for constructing 3D models from a single image. In this work, you will continue this novel work by further developing the tool toward complete automation similar to researchers in Stanford.

gui.png
  • C/C++ skills are required.

  • Supervisor: Prof Joni Kamarainen

  • Please contact the supervisor for more details.

-- JoniKaemaeraeinen - 31 Oct 2016
Topic attachments
I Attachment Action Size Date Who Comment
360_image.jpgjpg 360_image.jpg manage 107.6 K 21 Apr 2016 - 16:00 JoniKaemaeraeinen  
360_image_small.jpgjpg 360_image_small.jpg manage 28.4 K 21 Apr 2016 - 16:02 JoniKaemaeraeinen  
Lytro_ILLUM_DSC_7403-.jpgjpg Lytro_ILLUM_DSC_7403-.jpg manage 1030.0 K 08 Sep 2016 - 16:36 SaidPertuzarroyo Lytro Illum plenoptic camera
Raspberry_Pi_Logo_small.svg.pngpng Raspberry_Pi_Logo_small.svg.png manage 25.0 K 13 Jan 2016 - 10:31 JoniKaemaeraeinen  
big_data_flickr_infocux_small.jpgjpg big_data_flickr_infocux_small.jpg manage 48.4 K 29 Oct 2015 - 09:26 JoniKaemaeraeinen  
bscthesis_hiltunen_simple_features.pdfpdf bscthesis_hiltunen_simple_features.pdf manage 1505.8 K 06 Mar 2014 - 13:20 JoniKaemaeraeinen  
deepnn_yuan.pngpng deepnn_yuan.png manage 205.1 K 23 Nov 2015 - 15:39 JoniKaemaeraeinen  
dpm_detection.jpgjpg dpm_detection.jpg manage 66.8 K 21 Feb 2016 - 21:21 JoniKaemaeraeinen  
gui.pngpng gui.png manage 520.7 K 06 Mar 2014 - 12:12 JoniKaemaeraeinen Single image reconstruction GUI
hs_joni_magic_wand_small.jpgjpg hs_joni_magic_wand_small.jpg manage 28.3 K 19 Apr 2016 - 14:53 JoniKaemaeraeinen  
nostalgia_timeline_small.pngpng nostalgia_timeline_small.png manage 429.2 K 16 Oct 2015 - 09:45 JoniKaemaeraeinen  
pose_exmpl2.pngpng pose_exmpl2.png manage 86.1 K 19 Mar 2014 - 09:48 JoniKaemaeraeinen  
potters_wand.jpgjpg potters_wand.jpg manage 21.5 K 18 Sep 2015 - 08:53 JoniKaemaeraeinen  
stmapping.pngpng stmapping.png manage 361.7 K 08 Sep 2016 - 16:18 SaidPertuzarroyo Anatomical mammographic image analysis
visual_regression_small.pngpng visual_regression_small.png manage 92.0 K 16 Oct 2015 - 11:28 JoniKaemaeraeinen  
Print version |  PDF  | History: r25 < r24 < r23 < r22 | 
Topic revision: r25 - 31 Oct 2016 - 16:49:37 - JoniKaemaeraeinen
 

TUTWiki

Copyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TUTWiki? Send feedback