In this project we investigate the complex problem of object detection in images and videos. We are interested in studying and designing solutions that couple computer vision and learning ingredients and that can be specified to the problems at hand.
Our recent research focused on two classes of objects useful in video-surveillance scenarios (faces and pedestrians). In both cases we started of from over-complete dictionaries of image measurements and applied feature selection schemes to learn the most appropriate representation from data.
- Face detection: we deal with it according to a feature selection (see Fig. 1) and classification focus-of-attention pipeline . A data-driven approach allows us to exploit the peculiarities of different application domains (Fig. 2), and to capture the essence of the problem with respect to the application constraints. Multimedia: Face detection video advert (Augusto Destrero PhD dissertation, may 2008)
Figure 1: the first 42 rectangle features selected by our method as more representative for face detection
Figure 2: examples of results of the face detection procedure considering different settings
- Pedestrian detection: our objective for this work is the design of a solution able to automatically learn the representation from the training data and exploiting the features structure to improve the efficiency . We start off from a variable size HOG descriptor (even if our method can work with different features) and obtain a sparse representation of dictionary by means of Group Lasso, that allows us to preserve the block structure of HOG.
- Imavis (R&D),
- Istituto Chiossone (Usability, accessibility, testing for visually impaired users),
- Destrero, A. et al. "A Regularized Framework for Feature Selection in Face Detection and Authentication". IJCV (2009).
- Zini, Luca and Francesca Odone. "Efficient pedestrian detection with group lasso". Visual Surveillance - Workshop at ICCV, 2011. 1777-1784.
- Zini, L., A. Destrero and F. Odone. "A classification architecture based on connected components for text detection in unconstrained environments". Proc of AVSS, 2009.