Page d'accueil // SnT // Research // SIGCOM // Computer Vision Lab

Computer Vision Lab

The SnT Computer Vision Lab has been officially inaugurated on March 9, 2012. This lab provides state-of-the-art facilities for controlled imaging measurements, required for testing, and validation of image processing and computer vision algorithms. Activities of the SnT Computer Vision Lab are focused around real-world security applications such as video-based surveillance and automotive safety, and recently activities on automatic health monitoring have started. Dedicated dataset collected by the team and made available for the research community may be found here: SnT publicly available CV datasets.

Past and ongoing research and development activities in the SnT Computer Vision Lab are:

Multi-modal data fusion – Both fusion of different image modalities and fusion at different levels of abstraction are considered. Examples include fusion of 3D data captured with a low resolution depth sensor with 2D images of a high resolution camera, and multi-view fusion of RGB-D data. 

Depth data denoising and super-resolution – Enhancement of depth videos by developing denoising and super-resolution algorithms that take advantage of temporal information contained in the row videos. Special attention is given to videos containing dynamically deforming objects.

Activity recognition – The interest is to develop activity recognition methods that are robust to real-world conditions, where there is a constant change in illumination, texture, occlusions and viewpoint. Our focus is on using RGB-D cameras, from which 3D information can be coupled with colour information, thus reducing sensitivity to illuminations and textures. 

3D face recognition – Different phases of face recognition are considered; starting from 3D face reconstruction, feature extraction to classification. Face dynamics are also explored in the context of expression/emotion recognition.

RGB-D multi-view calibration and 3D reconstruction – In order to deal with occlusions, multiple RGB-D cameras at different viewpoints can be fused. To that end, the relative pose of each camera needs to be estimated while taking advantage of all the available data as captured by RGB-D cameras.   

2D/3D Shape modeling – The focus is on modeling non-rigid deformations of shapes in 2D and in 3D accounting for both their topological and geometric properties.    

Most Recent Projects

Dynamic Super Resolution of Depth Sequences with Non-Rigid Motions





RGB-D Multi-View System Calibration for Full 3D Scene Reconstruction





KinectDeform: Enhanced 3D Reconstruction of

Non-Rigidly Deforming Objects






Example-Dependent Cost-Sensitive Decision Trees






Template-based Statistical Shape Modelling On Deformation Space





COSTCLA: A Cost-Sensitive Classification Library in Python