Research Interests

• Robotics
• Computer Vision
• Machine Learning

Degrees

• PhD in Computer Science, Faculty of Electrical Engineering and Computing, University of Zagreb, 2023.
• MS in Mathematics and Computer Science, Department of Mathematics, J. J. Strossmayer University of Osijek, 2017.
• BS in Mathematics, Department of Mathematics, J. J. Strossmayer University of Osijek, 2015.

Publications

1. J. Maltar, I. Marković, I. Petrović, Visual Place Recognition using Directed Acyclic Graph Association Measures and Mutual Information-based Feature Selection, Robotics and Autonomous Systems 132 (2020)
   Abstract

   Visual localization is a challenging problem, especially over the long run, since places can exhibit significant variation due to dynamic environmental and seasonal changes. To tackle this problem, we propose a visual place recognition method based on directed acyclic graph matching and feature maps extracted from deep convolutional neural networks (DCNN). Furthermore, in order to find the best subset of DCNN feature maps with minimal redundancy, we propose to form probability distributions on image representation features and leverage the Jensen-Shannon divergence to rank features. We evaluate the proposed approach on two challenging public datasets, namely the Bonn and the Freiburg datasets, and compare it to the state-of-the-art methods. For image representations, we evaluated the following DCNN architectures: AlexNet, OverFeat, ResNet18 and ResNet50. Due to the proposed graph structure, we are able to account for any kind of correlations in image sequences, and therefore dub our approach NOSeqSLAM. Algorithms with and without feature selection were evaluated based on precision-recall curves, area under the curve score, best recall at 100% precision score and running time, with NOSeqSLAM outperforming the counterpart approaches. Furthermore, by formulating the mutual information-based feature selection specifically for visual place recognition and by selecting the feature percentile with the best score, all the algorithms, and not just NOSeqSLAM, exhibited enhanced performance with the reduced feature set.

1. D. Ševerdija, T. Prusina, A. Jovanović, L. Borozan, J. Maltar, D. Matijević, Compressing Sentence Representation with Maximum Coding Rate Reduction (Best paper award in AIS - Artificial Intelligence Systems track), ICT and Electronics Convention (MIPRO), 2023 46th MIPRO, Opatija, Hrvatska, 2023
   Abstract

   In most natural language inference problems, sentence representation is needed for semantic retrieval tasks. In recent years, pre-trained large language models have been quite effective for computing such representations. These models produce high-dimensional sentence embeddings. An evident performance gap between large and small models exists in practice. Hence, due to space and time hardware limitations, there is a need to attain comparable results when using the smaller model, which is usually a distilled version of the large language model. In this paper, we assess the model distillation of the sentence representation model Sentence-BERT by augmenting the pre-trained distilled model with a projection layer additionally learned on the Maximum Coding Rate Reduction (MCR2) objective, a novel approach developed for general purpose manifold clustering. We demonstrate that the new language model with reduced complexity and sentence embedding size can achieve comparable results on semantic retrieval benchmarks.
2. J. Maltar, D. Matijević, Optimization techniques for image representation in visual place recognition, 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO), Opatija, 2022, 877-882
   Abstract

   In visual place recognition we aim to match a given query image from a query database with the most appropriate reference image from a reference database. One of the main issues is how to represent a place. Although an ordinary RGB representation can represent a place, various, either handcrafted or learned representations such as deep convolutional neural networks achieve better quantitative results. By using optimization techniques, both convex and non-convex, we can adapt a place representation such that it fits into the problem of visual place recognition. Therefore, in this paper we examine numerous optimization techniques and incorporate them in the context of our problem. Quantitatively, in terms of the area under a curve (AUC) measure, conducted experiments show how such optimized representation outperforms unoptimized one.
3. J. Maltar, I. Marković, I. Petrović, NOSeqSLAM: Not only Sequential SLAM, Robot 2019: Fourth Iberian Robotics Conference, Porto, Portugal, 2019, 179-190
   Abstract

   The essential property that every autonomous system should have is the ability to localize itself, i.e., to reason about its location relative to measured landmarks and leverage this information to consistently estimate vehicle location through time. One approach to solving the localization problem is visual place recognition. Using only camera images, this approach has the following goal: during the second traversal through the environment, using only current images, find the match in the database that was created during a previously driven traversal of the same route. Besides the image representation method – in this paper we use feature maps extracted from the OverFeat architecture – for visual place recognition it is also paramount to perform the scene matching in a proper way. For autonomous vehicles and robots traversing through an environment, images are acquired sequentially, thus visual place recognition localization approaches use the structure of sequentiality to locally match image sequences to the database for higher accuracy. In this paper we propose a not only sequential approach to localization; specifically, instead of linearly searching for sequences, we construct a directed acyclic graph and search for any kind of sequences. We evaluated the proposed approach on a dataset consisting of varying environmental conditions and demonstrated that it outperforms the SeqSLAM approach.

Teaching

Courses:

• Web Programming
• Intelligent Robotics Systems
• Embedded Systems

Office Hours: Per request