Research Topics

Stereo Vision:

Our research in stereo vision aims at developing novel accurate stereo matching algorithms based on local and global approaches, as well as novel real-time stereo algorithms. Furthermore, we also research on stereo applications such as change detection algorithms for motion detection based on the 3D reconstruction provided by stereo vision.

• Accurate Stereo Vision
• Real-time stereo vision and applications (3D Intrusion Detection, 3D People Tracking / Counting)
• Classification and performance evaluation of different aggregation costs for stereo matching
• Unsupervised Adaptation for Deep Stereo (ICCV2017)
• Real-Time Self-Adaptive Deep Stereo

Surface Matching:

We develop novel techniques aiming at robust and efficient surface matching, which is commonly adopted for computer vision applications such as 3D Object Recognition and Surface Registration.

• SHOT: Unique Signatures of Histograms for Local Surface Description
• Repeatability of local reference frames for surface matching
• ReLOC: Pairwise registration by local orientation cues
• Pairwise 3D registration evaluation: methodology, datasets and results
• Learning a Descriptor-Specific 3D Keypoint Detector (ICCV15)
• Learning to Detect Good 3D Keypoints (IJCV18)

Domain Adaptation for Deep Learning:

We develop novel technique to address loose in performance due to domain shift between train and test data in deep learning.

• Exploiting Semantics in Adversarial Training for Image-Level Domain Adaptation (IPAS 2018)

Computer Vision for Grocery Stores management:

We are currently researching the best way to deploy state of the art computer vision algorithm to ease the management of supermarket and Grocery Stores as part of an industrially sponsored research.

• Product recognition in store shelves as a sub-graph isomorphism problem. (ICIAP17)
• A deep learning pipeline for product recognition on store shelves (IPAS18)

Joint Detection, Tracking and Mapping:

We are currently working on the problem of joint object detection and SLAM by means of a novel unified framework dubbed Semantic Bundle Adjustment. For more information and available code, please visit our project page.

Detection of repeatable keypoints in 3D data:

We are currently studying the problem of automatic extraction of repeatable and distinctive keypoints from 3D data. These keypoints can then be used for surface matching, by yielding point-to-point correspondences between a pair of point clouds. Usually, this is carried out by associating each keypoint with a compact representation of its local neighborhood (3D descriptor). 
We have carried out a survey and performance evaluation of the state of the art in the field of 3D keypoint detectors. For more information, including the results obtained and a link to the datasets used and to our papers, please visit our project page.

On-line Adaptive Video Tracking:

We develop novel techniques aiming at long-term video tracking, whereby on-line adaptation of the tracker behaviour is crucial. In particular, we have studied the problems of

• On-line Transition Model Adaptation
• On-line Appearance Model Adaptatio

Fast exhaustive Pattern Matching and Block Matching:

This research topic deals with developing novel techniques for fast pattern matching and block matching. We focus on exhaustive techniques, i.e. techniques that can accelerate the matching process and guarantee the same optimal solution as a full-search approach.

• Fast Exhaustive Pattern Matching and Block Matching

Detection of changes in video sequences:

This topic deals with extracting information on the observed scene from videos. The focus of the research is on robust change detection algorithms aimed at segmenting foreground objects from a reference background scene.

Image registration and mosaicing:

Registration is the process of determining the spatial and photometric relations among multiple views of a scene by exploiting correspondences within overlapping regions. Mosaicing is concerned with the fusion of several views into a unique image with augmented spatial and tonal extent and resolution.

• Automatic Image Mosaicing Algorithms
• Motion Detection with Pan-Tilt-Zoom Cameras using Background Mosaics
• Performance Evaluation Methodology for Image Mosaicing Algorithms

Augmented Reality:

Augmented Reality aims at supplementing reality by melting virtual objects and the real world. The virtual objects, i.e. 3D models, text, videos, are displayed as they were really there and at the right time. The augmented reality have to be both structurally and contextually synchronized with the real world and must perform independently of the user motion and the real environment.

Robust visual correspondence:

This research topic aims at developing novel measures for visual correspondence which are robust towards disturbance factors such as sudden illumination changes, noise, occlusions. So far, such measures have been usefully adopted for tasks such a pattern matching and change detection. Future work aims at estending their range of applications to stereo vision and image registration.

• Robust Visual Correspondence
• Robust Pattern Matching performance evaluation dataset
• Change detection robust to sudden illumination changes

RGB-D datasets and experimental evaluation:

Our work on RGB-D Visual Search has led to the realization of the HyperRGBD framework, a set of tools conceived to enable seamless aggregation of existing RGB-D datasets in order to obtain new data featuring desired peculiarities and cardinality. 

Industrials Applications:

We cooperate with several companies in order to develop advanced machine vision applications. Currently we are addressing the industrial safety application domain, quality control for the ceramics industry, vision-based access control for high security gates and computer vision for retails. These cooperations involve respectively DATASENSOR, SYSTEM GROUP, PLEXA. and Centro Studi SRL

Youtube Channel

Follow our work also on Youtube through our Youtube channel.