Publications by Rita Cucchiara

On-line action recognition from a continuous stream of actionsis still an open problem with fewer solutions proposedcompared to time-segmented action recognition. The mostchallenging task is to classify the current action while findingits time boundaries at the same time. In this paper wepropose an approach capable of performing on-line actionsegmentation and recognition by means of batteries of HMMtaking into account all the possible time boundaries and actionclasses. A suitable Bayesian normalization is appliedto make observation sequences of different length comparableand computational optimizations are introduce to achievereal-time performances. Results on a well known actiondataset prove the efficacy of the proposed method

In this paper a document analysis tool for historical manuscripts is proposed. The goal is to automatically segment layout components of the page, that is text, pictures and decorations. We specifically focused on the pictures, proposing a set of visual features able to identify significant pictures and separating them from all the floral and abstract decorations. The analysis is performed by blocks using a limited set of color and texture features, including a new texture descriptor particularly effective for this task, namely Gradient Spatial Dependency Matrix. The feature vectors are processed by an embedding procedure which allows increased performance in later SVM classification.

This paper proposes a comparison of color features for image retrieval. In particular the UCID image database has been employed to compare the retrieval capabilities of different color descriptors. The set of descriptors comprises global and spatially related features, and the tests show that HSV based global features provide the best performance at varying brightness and contrast settings.

In this paper we present an overview of the historical evolution of connected component labeling algorithms, and in particular the ones applied on images stored in raster scan order. This brief survey aims at providing a comprehensive comparison of their performance on modern architectures, since the high availability of memory and the presence of caches make some solutions more suitable and fast. Moreover we propose a new strategy for label propagation based on a 2x2 blocks, which allows to improve the performance of many existing algorithms. The tests are conducted on high resolution images obtained from digitized historical manuscripts and a set of transformations is applied in order to show the algorithms behavior at different image resolutions and with a varying number of labels.

The detection of humans in very complex scenes can be very challenging, due to the performance degradation of classical motion detection and tracking approaches. An alternative approach is the detection of human-like patterns over the whole image. The present paper follows this line by extending Tuzel et al.’s technique [1] based on covariance descriptors and LogitBoost algorithm applied over Riemannian manifolds. Our proposal represents a significant extension of it by: (a) exploiting motion information to focus the attention over areas in which motion is present or was present in the recent past; (b) enriching the human classifier by additional, dedicated cascades trained on positive and negative samples taken from the specific scene; (c) using a rough estimation of the scene perspective, to reduce false detections and improve system performance. This approach is suitable in multi-camera scenarios, since the monolithic block for human-detection remains the same for the whole system, whereas the parameter tuning and set-up of the three proposed extensions (the only camera-dependent parts of the system), are automatically computed for each camera. The approach has been tested on a construction working site in which complexity and dynamics are very high, making human detection a real challenge. The experimental results demonstrate the improvements achieved by the proposed approach.

This article presents a framework for automatic semantic annotation of video streams with an ontology that includes concepts expressed using linguistic terms and visual data.

In this paper we present a new optimization technique for the neighborhood computation in connected component labeling focused on images stored in raster scan order. This new technique is based on a 2x2 square block analysis of the image, and it exploits the fact that, when using 8-connection, the pixels of a 2x2 square are all connected to each other. This implies that they will share the same label at the end of the computation. To prove the effectiveness of our proposal, we show a comprehensive comparison of the most used and advanced connected components labeling techniques presented so far. The tests are conducted on high resolution images obtained from digitized historical manuscripts and a set of transformations is applied in order to show the algorithms behavior at different image resolutions and with a varying number of labels.

This paper proposes a system for people trajectory shape analysis by exploiting a statistical approach which accounts for sequences of both directional (the directions of the trajectory) and linear (the speeds) data. A semi-directional distribution (AWLG - Approximated Wrapped and Linear Gaussian) is used with a mixture to find main directions and speeds. A variational version of the mutual information criterion is proposed to prove the statistical dependency of the data. Then, in order to compare data sequences, we define an inexact method with a Kullback-Leibler-based distance measure and employ a global alignment technique is to handle sequences of different lengths and with local shifts or deformations. A comprehensive analysis of variable dependency and parameter estimation techniques are reported and evaluated on both synthetic and real data sets.

This paper presents a novel approach for detecting multiple instances of the same object for pick-and-place automation. The working conditions are very challenging, with complex objects, arranged at random in the scene, and heavily occluded. This approach exploits SIFT to obtain a set of correspondences between the object model and the current image. In order to segment the multiple instances of the object, the correspondences are clustered among the objects using a voting scheme which determines the best estimate of the object's center through mean shift. This procedure is compared in terms of accuracy with existing homography-based solutions which make use of RANSAC to eliminate outliers in the homography estimation.