About

Domen Tabernik is a research associate at the Visual Cognitive Systems (ViCoS) Laboratory at the University of Ljubljana, Faculty of Computer and Information Science. His doctoral dissertation investigated hierarchical and compositional representations for visual recognition using deep neural models. His current research focuses on deep learning methods for computer vision, particularly visual anomaly detection and industrial surface inspection. He develops data-efficient models for defect detection and segmentation in industrial environments, including approaches based on weak or mixed supervision that reduce the need for extensive manual annotation while maintaining high detection accuracy. Beyond industrial inspection, his work includes point-supervised learning methods for object counting and localization, such as dense center-direction regression models, as well as vision-based methods for robotic manipulation, including grasp detection and handling of deformable objects. He has contributed to several research and development projects on data-driven machine vision systems, advancing the practical deployment of deep learning in industrial inspection, infrastructure monitoring, and robotics.

Research Topics

Center-Directions for counting and localization

This research focuses on developing a novel point-supervised learning approach for object counting and localization based on regression of center-directions. We introduce an architecture for dense regression of center-directions, splitting the problem into domain-specific dense regression and a lightweight, domain-agnostic localization network.

Vision for robotic manipulation

Developing novel vision-based methods for robotic manipulation, with a focus on grasping deformable objects such as cloths, towels, and garments. Introducing CeDiRNet-3DoF and the ViCoS Towel Dataset for benchmarking and advancing cloth manipulation in robotics.

Industrial surface defect detection

The developed methods allow specialization for large defect detection on various indistrual items such as cracks, smudges, imperfections etc. The methods are learning-based and are thus robust, run realtime and are applicable to a wide range of real problems.

Traffic-sign detection

We explore automation of traffic-sign inventory management using deep-learning models. Models such as Faster R-CNN and Mask R-CNN are improved and applied to traffic sign detection. Instead of specializing in automated detection for only several traffic sign categories we explore possibility of automating the detection of over 200 different traffic signs that are needed to automate the traffic-sign inventory management.

Deep compositional networks

We propose a novel deep network architecture that combines the benefits of discriminative deep learning and the benefits of compositional hierarchies. As one of the benefits we emphasize the ability to automatically adjust receptive fields to either small or large receptive fields depending on the for problem at hand and the ability to visualize deep features through explicit compositional structure

Learning a hierarchy of parts

We deal with a problem of Multi-class Object Representation and present a framework for learning a hierarchical shape vocabulary capable of representing objects in hierarchical manner using a statistically important compositional shapes. The approach takes simple oriented contour fragments and learns their frequent spatial configurations. These are recursively combined into increasingly more complex and class specific shape compositions, each exerting a high degree of shape variability

Histogram of compositions

As extension to LHOP model, we have developed a shape descriptor capable of using compositional parts learnt using the LHOP model to provide a descriptor that is compatible with HOG descriptor and can be easily used as direct replacement.

ViCoS Eye

ViCoS Eye is an experimental online service that aims to demonstrate a state-of-the-art computer vision object detection and categorization algorithm developed in our laboratory. Web-service is available in a form of a web-page and in a form of an Android application.

Downloads and Code

Current projects

EOFuseR - Earth Observation with Sensor-Fusion and Representation Learning

January 2025 - April 2026

This ESA funded project investigates the relationship between sensor fusion and self-supervised learning for data-driven Earth Observation. We focus on the role of self-supervised deep learning for sensor fusion from the perspective of different sources with different spatial resolutions and spectral coverage. The project is grounded in a real-world application in the field of hydrology, where the goal is to predict the water level in rivers using satellite and drone imagery.

RoDEO - Robust Deep Learning for Earth Observation​

January 2025 - December 2027

This ARIS funded project investigats the relationship between sensor fusion and self-supervised learning for data-driven Earth Observation. We focus on the role of self-supervised deep learning for sensor fusion from the perspective of different sources with different spatial resolutions and spectral coverage. The project is grounded in a real-world application in the field of hydrology, where the goal is to predict the water level in rivers using satellite and drone imagery.

MUXAD - Multimodal Image Understanding for Explainable Anomaly Detection

January 2025 - December 2027

The functional objective of the project is to advance anomaly detection in images by developing multimodal models that integrate visual and linguistic information to not only detect if and where anomalies occur but also explain why. The main research goal is to create novel methods for semantic image understanding, zero-shot multimodal anomaly detection, and multimodal explanations that improve AI’s interpretability and transparency while reducing reliance on annotated data.

Computer Vision

January 2019 - December 2027

Computer vision is becoming a focal problem area of artificial intelligence. On the wings of deep learning it has become very powerful tool for solving various problems involving processing of visual information. In the framework of this programme we are addressing several research questions ranging from visual tracking to visual learning for autonomous robots, with a special emphasis on going beyond supervised deep learning.

Past projects

MV4.0 - Data-driven framework for development of machine vision solutions

October 2021 - September 2024

The functional objective of the project is to shift the paradigm in the development of machine vision solutions from hand-engineered specific solutions to data-driven learning-based design and development that would enable more general, efficient, flexible and economical development, deployment and maintenance of machine vision systems. The main research goal of this project is to develop novel deep learning methods for iterative, active, robust, weak, self-, unsupervised and few-shot learning that would reduce the amount of needed annotated data.

DIVID - Detection of inconsistencies in complex visual data using deep learning

July 2018 - December 2021

The objective of the project is to develop novel deep learning methods for modelling complex consistency and detecting inconsistencies in visual data using training images annotated with different levels of accuracy. The main project goal is to go beyond the traditional supervised learning, where all anomalies on all training images have to be adequately labelled.

GOSTOP - Building Blocks, Tools and Systems for the Factories of the Future

November 2016 - January 2020

The aim of the GOSTOP programme was to accelerate the development of the Factories of the Future concept in Slovenia and to provide solutions to the current needs of Slovene industry. Our goal was to develop efficient machine vision algorithms, coupled with machine learning approaches, which would allow for fast and flexible adaptation of visual inspection systems to be able to deal with novel quality control problems.

ViLLarD - Maintenance of large databases based on visual information using incremental learning

July 2014 - June 2017

The main goal of the project is to develop a framework for semi-supervised interactive incremental learning as well as specific methods for visual learning and recognition that will increase the quality and efficiency of large visual information databases maintenance.

CV4foot - Study and comparison of advanced computer vision methods for foot modelling in a real-world environment

April 2014 - September 2014

In this student project we were exploring the potential of using computer vision techniques for footwear recommendation systems. The maingoal was to improve existing methods with advanced computer vision technologies, to solve the problem of automatic feet modelling, and to determine the suitability of the latest mobile devices for such advanced computer vision algorithms.

LeOParts - Learning a large number of visual object categories for content-based retrieval in image and video databases

April 2010 - August 2013

The challenge this project addressed was development of a methodology that would bridge the gap between the computer-centered low-level image features and the high-level human-centered semantic meanings. The methodology explored was hierarchical compositional models, enriched by discriminative information and extended to online learning.

Computer vision for mobile computing and interaction (RS)

January 2009 - December 2012

The use of computer vision makes for a very intuitive interaction with mobile device, greatly simplifying it. We developed computer vision methods suitable for mobile devices, and use them to implement designed interaction scenarios in prototype applications.

Publications