Neven Pičuljan

Created a feasibility study and demo for a face recognition system that I developed at Visage Technologies. The demo was written in C/C++ and Python. I collaborated with the Swedish Police, the Swedish National Forensic Center, and the Swedish Defence Research Agency on this project.

A neural network for function approximation using the Levenberg-Marquardt algorithm. I tested the code on various functions and used Torch framework and Python.

A simple classifier in R for credit card applications. I was given a data set with users’ interactions and experimented with various machine learning algorithms: SVMs, decision trees, random forests, logistic regression, etc.

A project in data mining. I was given a data set with detailed information about interactions of visitors with different stations at Copernicus Science Centre in Warsaw, Poland. The goal was to characterize the flow of visitors through these stations and to segment visitors into separate categories/segments. I used R.

A system for detection and localization of a 2D contour (human head) in an image, where many such contours of different size could exist. For this purpose, I applied the generalized Hough Transform (GHT). The system was written in Python.

An implementation of various operations on graphs in LISP: finding cycles in graphs, finding paths from one node to another in graphs, checking if the binary tree is symmetric, depth-first order graph traversal, finding maximum depth of a binary tree, and finding a leaf with a maximum value in a binary tree.

An expert system created in Prolog for animal identification.

A face recognition system. It was trained and tested in Torch framework. The data set was made of publicly available data sets.

Research conducted on different algorithms for face alignment.

A system to automate the classification of Stack Overflow's posts in the answer thread into three categories: One for those that answer the question well. Another for those that can be potentially useful to the user (e.g., because they can help educate him/her on the subject). Lastly, group those that are just bad or useless.

I experimented with various machine learning algorithms (scikit-learn): Gaussian naive Bayes, SVMs, and random forests.

A system to do pedestrian detection in urban environments using contour based detection. It was written in Python using Numpy, Scikit-learn, and OpenCV.

Performance-driven animation as a web application. Face tracking was used to gain motion in the face of an animated virtual character. The graphics system used to build the application was Three.js based on WebGL. The face-tracking system used to build the application was Visage|SDK.

A computer game controlled with head movements. It was written in C# using Unity game engine.

Doctoral thesis

Object detection is a computer vision technique that enables machines to recognize, identify, and locate objects in an image. The dissertation aims to automate the quality assurance of the object bounding box annotations on images to create a dataset that can be used in object detection solutions in computer vision. In the object detection context, labels are a means of identifying and marking objects in images. They are defined by the object’s name and two points in the image coordinate.

URL: https://repozitorij.fer.unizg.hr/islandora/object/fer:13478

Pičuljan, N., & Končar, M. (2023, November). Secure mobile access to patient imaging data using SMART on FHIR. In Medicinska informatika 2023 – 16. simpozij Hrvatskog društva za medicinsku informatiku (pp. 66–70). Hrvatsko društvo za medicinsku informatiku

Pičuljan, N., & Car, Ž. (2022, September). Semantic Metadata for Image Files. In 2022 International Symposium ELMAR (pp. 203–208). IEEE.

Deep neural networks used in industry applications usually work the best when they are trained using supervised learning given that there is a lot of data available, that the data is from the same distribution as the data from the production environment and given that the labels are of a high quality. Large amounts of image data is available on the Internet, but it is not useful for machine learning applications in raw format. This concept goes through the process of image data acquisition from public sources on the Internet and automatic data labeling using existing computer vision deep learning models for object detection and instance segmentation. It sets up baseline labels that way and touches the problem of versioning labels from different sources for the same image and the same task. Eventually, it assigns machine learning semantic metadata to the given image. Furthermore, it goes beyond baseline automatic labels and proposes concept to refine them in both automatic and manual ways. With the semantic metadata available, it is shown how to make use of that semantic metadata to synthetica...

Pičuljan, N., & Car, Ž. (2023). Machine Learning-Based Label Quality Assurance for Object Detection Projects in Requirements Engineering. Applied Sciences, 13 (10), 6234.

In recent years, the field of artificial intelligence has experienced significant growth, which has been primarily attributed to advancements in hardware and the efficient training of deep neural networks on graphics processing units. The development of high-quality artificial intelligence solutions necessitates a strong emphasis on data-centric approaches that involve the collection, labeling, and quality assurance of data and labels. These processes, however, are labor-intensive and often demand extensive human effort. Simultaneously, there exists an abundance of untapped data that could potentially be utilized to train models capable of addressing complex problems. These raw data, nevertheless, require refinement to become suitable for machine learning training. This study concentrates on the computer vision subdomain within artificial intelligence and explores data requirements within the context of requirements engineering. Among the various data requirement activities, label quality assurance is crucial.

URL: https://www.mdpi.com/2076-3417/13/10/6234

Heidenreich, H., Dalvi, R., Mukku, R., Verma, N., Pičuljan, N. (2024). Large Language Models for Page Stream Segmentation. arXiv preprint arXiv:2408.11981.

Page Stream Segmentation (PSS) is an essential prerequisite for automated document processing at scale. However, research progress has been limited by the absence of realistic public benchmarks. This paper works towards addressing this gap by introducing TABME++, an enhanced benchmark featuring commercial Optical Character Recognition (OCR) annotations. We evaluate the performance of large language models (LLMs) on PSS, focusing on decoder-based models fine-tuned with parameter-efficient methods. Our results show that decoder-based LLMs outperform smaller multimodal encoders. Through a review of existing PSS research and datasets, we identify key challenges and advancements in the field. Our findings highlight the key importance of robust OCR, providing valuable insights for the development of more effective document processing systems.

The project focuses on building an intelligent pricing engine that uses computer vision and natural language processing to analyze product images, descriptions, and metadata, helping generate accurate and market-aware price recommendations.