Computer Vision

The research program group is involved in fundamental and applicative research in computer and cognitive vision. This area is one of the most important components of intelligent systems and represents a crucial part of numerous applications such as detection, categorisation and segmentation of objects and scenes, automatic query in large pictorial and video databases, face recognition, analysis of human behaviour, visual surveillance and tracking, autonomous driving (cars, aerial vehicles, vessels) and controlling of different types of robots. In the last few years, riding the success of deep learning, computer vision made a big leap forward. Practical applications of computer vision are moving out of constrained industrial environments into different domains of every day life. Still, many problems remain. Robust operation in dynamic and complex environment is still not sufficient for complete autonomy. For successful operation of deep learning methods, huge amounts of annotated learning data is required, making this approach limited and expensive. Fortunately, access to large amounts of images, video and 3D point data is getting simpler and cheaper. Computing power necessary for processing is also increasing. In the framework of this research program we will address the current limitations and use the progress of deep learning to our advantage. We will perform fundamental research and apply these findings on specific problem domains. We will continue our successful work in the area of visual tracking and methodology for their evaluation. We will develop new methods of computer vision and machine learning for autonomous control of mobile robots. We will introduce the concept of compositional deep models in the area of deep learning and introduce techniques that overcome the classical supervised discriminative approach. We will use deep learning in different problem domains such as segmentation and modelling of 3D point clouds and biometry. We will also develop new user interfaces using cameras. The program will, as up till now, devote much of its activities to evaluation of theoretical results on various real platforms, such as mobile robots, active sensory systems and smart mobile devices. We will invest our efforts into building publicly accessible annotated image and video databases, organising challenges in the framework of international conferences and proposing suitable evaluation protocols and metrics. Finally, our research program will transfer theoretical knowledge to practical applications, also in cooperation with the end users, building on existing practise (use of computer vision for web sales, documentation of cultural heritage, participation in new media art, …).

The main purpose of computer vision is to automatically derive information from images or from a sequence of images such as video. Some typical tasks that fit into computer vision are identification of people, based on images of their faces, tracking of an object in a video sequence, building of 3D models of individual objects or a complete environment from a set of images, and analysis or examination of images according to some medical or even aesthetic  criteria. With development of IT, capture and sharing of visual information is much simplified and substantial, and therefore the importance and the share of visual information in the entire information space is much larger. For example, in almost all social networks visual information predominates. Even older visual information from archives are being digitised in can be searched and analysed using computer vision. Routine manual searches for specific content in hundreds of hours of video is simply an unsurmountable task. Therefore computer vision is becoming a key method for managing, searching, classification and analysis of information in general. New  application areas are getting into focus, such as autonomous driving, where perception of visual information about the road, of other traffic participants, and of obstacles on the road are crucial. Finally, computer vision is a critical technology for several other technical, natural and  social sciences where automatic analysis of visual information can help in discoveries. Even humanities can benefit from computer vision attested by the advent of digital humanities. Analysis of historic artefacts in museums, analysis of animal or human behaviour, detecting changes on the earth’s surface from satellite imagery can all be done by computer vision. From a methodological and technical point of view, computer vision methods  went through radical changes in the last few years. In the past, different computer vision tasks demanded a high degree of specialisation, in particular, in determining image features needed for s particular task. Today convolutional neural networks are in predominance, since they can determine and search for relevant features in the input images themselves, based on their previous training on similar visual data.

Use of computer vision methods in actual use in practice requires a lot of experience since just applying some standard software solutions is not sufficient for demanding problem domains. Scientific activity in computer vision therefore serves also for human resources development for the most advanced automatic analysis of visual data. Without experts for computer vision several other fields that rely on computer vision would be set back. These fields range from robotics, which is traditionally tightly connected to computer vision, but also for analysis of medical images to offer faster and better medical interpretation. Automatic interpretation of video surveillance data is getting more and more important, and finally even automatic analysis of video shared over social networks is becoming imperative. Slovenian scientists are contributing to advances in computer vision in international framework, attested also by the high number of citations. The publication of our research program have almost 15.000 citations on Google Scholar.