semantic image labelling

Physical objects are the units in the real world to which people associate basic knowledge about the environment. The ultimate goal of semantic image labelling is to detect and classify all the objects present in an image. For its importance and challenge, automatic semantic labelling of images is an area that attracts many research interests.
In the past we have developed algorithms to describe an image in terms of low level features such as color, edges, texture. The results have been used to index image archives for a subsequent retrieval by similarity. Analogous algorithms are typically tailored to segment images in multiple regions: this often is the first step towards the detection of objects in specific applications. An orthogonal approach relies on the well known template matching techniques. In both cases a model of the object should be provided for a proper recognition.

activities and projects

• MEMORI - MEMory-based Object Recognition in Images (memory means a database collecting all the objects of interest depicted from different point of views which is exploited for the labelling of the scene).
• TM - Template Matching Techniques in Computer Vision: Theory and Practice. Look inside through Google Book - Template Matching Techniques in Computer Vision: the Code Companion [here]
• Visual Environmental Monitoring - Mountains etc. labelling by making a correspondence between peak profiles and 3D model of the Earth.
• COPILOSK - COntent Processing by Integrating LOgical and Statistical Knowledge (FBK joint research project) - In the framework of semantic image labelling the idea is to combine image analysis with knowledge-assisted techniques that use ontologies and domain knowledge (in collaboration with DKM research unit) to aggregate regions in the typically oversegmented image and to assign semantic labels to these regions.

dynamic scene understanding

Understanding a dynamic visual scene is the core problem of computer vision. Its goal is the detection and recognition of spatio-temporal patterns in video steams. For this reason modeling events is one of the key issue to describe the dynamic structure of a visual scene. Spatial and temporal knowledge along with specific knowledge and measured parameters of a changing environment must be considered.

activities and projects

• SmarTrack - a SmarT people Tracker. It is an adaptive system for real-time tracking of multiple people in a monitored scene. It provides accurate information about the spatial location of people through multiple persistent occlusions in cluttered environment using a number of timestamped image streams.
• TRAVEL - Traffic Road Analysis by Visual Event Labelling. Since several years TeV is involved in automatic analysis of traffic sequences from static or moving cameras. A flexible system for counting, classifying and tracking vehicles in road intersections is SCOCA.
• MY-e-DIRECTOR 2012 - Real-Time Context-Aware and Personalized Media Streaming Environments for Large Scale Broadcasting Applications (EU Project). The main goal of this project is to research and develop a solution that will allow the viewers to select focal actors and points of interest within real-time broadcasted scenes such as the Olympic Games of London 2012.
• NETCARITY - A NETworked multisensor system for elderly people: health CARe, safety and securITY in home environment (EU Project). The project investigates how new and existing technologies can be integrated cost effectively into elderly people's homes, making them feel more comfortable about remaining in their familiar environment improving their wellbeing, independence, and safety.
• ACube - Ambient Aware Assistance (PAT Grand Project). The goal is to improve the quality of life for the elderly and disabled through technological progress in areas such as a rehabilitation center or an institution for Alzheimer sufferers. ACube acts in the framework of Ambient Intelligence: the scientific challenge is to study methods and technologies for monitoring complex environments using distributed sensor networks in order to detect events, situations, and activities even in complex scenarios with many people.
• PUMALab - attentive Perception with a Unified approach to Multi-modality, Adaptation and Learning (FBK joint research project) (some examples)

activity and projects in the past

contact details

For information please contact
Stefano Messelodi  |  e-mail   m e s s e l o d (at) fbk . eu