Events

Defence of doctoral thesis in the field of computer science, Zakaria Laskar

Deep Learning Methods for Image Matching
CS_defence photo by Matti Ahlgren

Title of the doctoral thesis is "Deep Learning Methods for Semantic Matching, Image Retrieval and Camera Relocalization"

From robotics to augmented reality applications, the problem of image matching is a fundamental problem in computer vision. Verifying whether two images are similar or not, and if similar which parts are in correspondence can provide sufficient cues to address a wide range of autonomous scene understanding problems. For example, creating a 3D map of an environment from captured images, estimating camera motion for robust and accurate robot localization or interactive augmented reality applications etc. With the availability of large amount of data, image matching methods have drifted towards data-driven methods based on deep learning and convolutional neural networks. This opened both possibilities and challenges as data-driven methods suffer from the drawbacks of computational and memory inefficiency.

In this thesis, we present methods that address the problem of efficiency in several important directions. Deep learning methods require large amount of labelled data that requires excessive amount of human labor and proves to be inefficient for the problem of image matching. To address this, we propose methods that allows deep learning methods to learn from small amount of training data. Furthermore, the good performance of deep learning methods is generally attributed to its large size which has a high memory footprint in computer hard-drives. This is addressed in two prominent ways. Firstly, we propose methods to reduce the number of parameters in deep learning based image matching model that effectively reduces the memory size of the model. Secondly, a generic training algorithm is proposed that allows a single deep learning model to be deployed across several environments, thereby increasing memory efficiency by reducing the requirement of storing multiple models in memory.

Opponent: Professor Giorgios Tolias, Czech Technical University in Prague, Czech Republic

Custos: Professor Juho Kannala, Aalto University School of Science, Department of Computer Science

Doctoral candidate's contact information: Zakaria Laskar, Aalto University School of Science, Department of Computer Science, [email protected], +358449143242

The defence will be organized via remote technology (Zoom). Link to the defence.

Zoom Quick Guide (www.aalto.fi)

The doctoral thesis is publicly displayed 10 days before the defence in the publication archive of Aalto University (aaltodoc.aalto.fi).

Electronic dissertation (aaltodoc.aalto.fi)

  • Published:
  • Updated:
Share
URL copied!