TeV's research contributes to the mission of the Digital Industry Center, in that it carries out projects in various digital industry fields, investigating specific areas to achieve scientific results and developing new competencies. Basic research in computer vision helps expand existing knowledge and creates a foundation for solving future problems. Cutting-edge topics of Computer Vision have various names, but the ultimate goal is the description of visual inputs.

In 1966, Prof. Marvin Minsky hired a first-year undergraduate student, Gerald Sussman, and assigned him a problem to solve over the summer: “Connect a camera to a computer and get the machine to describe what it sees.” That summer, for our scientific community, is still going on. Relaxing more and more the assumptions about the knowledge of the world and the constrains on devices, leads us and the whole Computer Vision community in the study of algorithms to solve the task set by Minsky, to describe a (dynamic) scene geometrically starting from images or videos freely acquired by several cameras, possible moving in space, and depicting many objects in motion.

Fortunately, some specific problems attached with Computer Vision do not require the global description of a scene, but their solution benefits from the development and application of specific techniques and technologies.

We are currently actively working on the following research topics.


logo FBK Digital Industry Center
Logo Technologies of Vision Research Unit

machine vision

Machine Vision is comprised of scenarios in which the operational guidance of other devices is based on knowledge extracted from images. We approach Machine Vision applications with traditional image analysis, machine learning techniques or hybrid approaches, for example for robotic manipulation.

geometric deep learning

Geometric deep learning is a field of machine learning in which the machine learns from complex data represented through graphs or multi-dimensional points, possibly in non-euclidean spaces. Applications of geometric deep learning in computer vision include shape analysis and articulated object pose recognition.

object X

where X means: localization, detection, segmentation, classification, recognition, reconstruction, tracking. These are basic steps toward the semantic description of visual data. We are developing innovative algorithms in 6d object pose estimation, fine-grained classification, and re-identification. Furthermore, we are studying techniques to generate correspondences between images depicting the same object to provide a geometric reconstruction of that specific piece of the world and/or to track its motion.

event and action recognition

An event is a basic unit of information in streaming data. The detection and classification of events is useful for generating video descriptions and video retrieval, as well as for activity recognition in video surveilled environments for example to detect anomalies.

CV for extended reality

Human-machine interactions, augmented and mixed reality applications benefit from scene understanding modules to properly embed information and/or virtual objects into artificially generated scene. Extended reality holds great potential in industrial applications.

low level image processing

In the early stages of perception, visual analysis is complicated by different appearances of the same item, due to factors like: illumination, type of camera, presence of other objects, possibility of occlusions, motion, etc.. To tackle these problems, low-level analysis (or early vision) does not focus on semantic objects but instead on the scene as a whole. With the development and application of traditional and machine learning approaches, we perform research into Retinex theory, image enhancement, background updating, SLAM, and motion description. Furthermore we assign perceptive attributes to image using deep learning classifications.

visual pattern recognition

Pattern recognition has many applications in image analysis, with some of the classical tasks being OCR and fingerprint analysis. Landmark detection in scenes is an example of object/pattern recognition for self calibration, self localization and navigation. Anomalous pattern detection can be applied to industrial visual inspection for defect or damage identification.

disclaimer: images are taken from TeV research seminars

What about machine learning in out unit? Not a new topic to us. The partial renunciation of complicated elaborations in favour of the constitution of large databases, acquired through experience and associated with appropriate commands, seems to be the best way to simulate and realise what we commonly consider an intelligent activity. (Luigi Stringa, director, 1993)