Riccardo Satta

A multiple component matching framework for person re-identification

Person re-identification consists in recognizing an individual that has already been observed over a network of cameras. It is a novel and challenging research topic in computer vision, for which no reference framework exists yet. Despite this, previous works share similar representations of human body based on part decomposition and the implicit concept of multiple instances. Building on these similarities, we propose a Multiple Component Matching (MCM) framework for the person reidentification problem, which is inspired by Multiple Component Learning, a framework recently proposed for object detection [3]. We show that previous techniques for person re-identification can be considered particular implementations of our MCM framework. We then present a novel person re-identification technique as a direct, simple implementation of our framework, focused in particular on robustness to varying lighting conditions, and show that it can attain state of the art performances.

Fast person re-identification based on dissimilarity representations

Person re-identification is a recently introduced computer vision task that consists of recognising an individual who was previously observed over a video-surveillance camera network. Among the open problems, in this paper we focus on computational complexity. Despite its practical relevance, especially in real-time applications, this issue has been overlooked in the literature so far. In this paper, we address it by exploiting a framework we proposed in a previous work. It allows us to turn any person re-identification method, that uses multiple components and a body part subdivision model, into a dissimilarity-based one. Each individual is represented as a vector of dissimilarity values to a set of visual prototypes, that are drawn from the original non-dissimilarity representation. Experiments on two benchmark datasets provide evidence that a dissimilarity representation provides very fast re-identification methods. We also show that, even if the re-identification accuracy can be lower (especially when the number of candidates is low), the trade-off between processing time and accuracy can nevertheless be advantageous, in real-time application scenarios involving a human operator.

Multimodal Person Reidentification Using RGB-D Cameras

Person reidentification consists of recognizing individuals across different sensors of a camera network. Whereas clothing appearance cues are widely used, other modalities could be exploited as additional information sources, like anthropometric measures and gait. In this paper, we investigate whether the reidentification accuracy of clothing appearance descriptors can be improved by fusing them with anthropometric measures extracted from depth data, using RGB-D sensors, in unconstrained settings. We also propose a dissimilarity-based framework for building and fusing the multimodal descriptors of pedestrian images for reidentification tasks, as an alternative to the widely used score-level fusion. The experimental evaluation is carried out on two data sets including RGB-D data, one of which is a novel publicly available data set that we acquired using Kinect sensors. The fusion with anthropometric measures increases the first-rank recognition rate of clothing appearance descriptors up to 20%, whereas our fusion approach reduces the processing cost of the matching phase.

Exploiting dissimilarity representations for person re-identification

Person re-identification is the task of recognizing an individual that has already been observed over a network of videosurveillance cameras. Methods proposed in literature so far addressed this issue as a classical matching problem: a descriptor is built directly from the view of the person, and a similarity measure between descriptors is defined accordingly. In this work, we propose a general dissimilarity framework for person re-identification, aimed at transposing a generic method for person re-identification based to the commonly adopted multiple instance representation, into a dissimilarity form. Individuals are thus represented by means of dissimilarity values, in respect to common prototypes. Dissimilarity representations carry appealing advantages, in particular the compactness of the resulting descriptor, and the extremely low time required to match two descriptors. Moreover, a dissimilarity representation enables various new applications, some of which are depicted in the paper. An experimental evaluation of the proposed framework applied to an existing method is provided, which clearly shows the advantages of dissimilarity representations in the context of person re-identification.

Sensor Pattern Noise Matching Based on Reliability Map for Source Camera Identification

Source camera identification using the residual noise pattern left by the sensor, or Sensor Pattern Noise, has received much attention by the digital image forensics community in recent years. One notable issue in this regard is that high-frequency components of an image (textures, edges) can be easily mistaken as being part of the SPN itself, due to the procedure used to extract SPN, which is based on adaptive low-pass filtering. In this paper, a method to cope with this problem is presented, which estimates a SPN reliability map associating a degree of reliability to each pixel, based on the amount of high-frequency content in its neighbourhood. The reliability map is then used to weight SPN pixels during matching. The technique is tested using a data set of images coming from 27 different cameras; results show a notable improvement with respect to standard, non-weighted matching.

Sensor pattern noise and image similarity for picture-to-identity linking

Picture sharing through social networks has become a prominent phenomenon, producing a large amount of data that law enforcers may be entitled to use, under the proper legal framework, as a source of information for investigating a crime. In this work, the authors exploit digital camera ‘fingerprinting’ based on noise residuals (sensor pattern noise or SPN) to achieve a novel forensic task, named picture-to-identity linking. It consists of finding social network accounts that possibly belong to the author of a certain photo (e.g. showing illegal content). The rationale is that the author of the offending photo has likely used the same camera for taking other (legal) pictures, and posted them in a social network account. The authors extend a previous work on the topic by coupling SPN with visual image similarity, a useful cue when pictures have been taken in the same environment (e.g. a room). The authors also improve the framework by allowing for multiple-image queries, and thoroughly evaluate the performance on two corpora of images from social network accounts, including the impact of image modifications. Reported results show a robust improvement with respect to the previous work, and prove the usefulness of picture-to-identity as an aid for digital forensic investigations.

Exploiting depth information for indoor-outdoor scene classification

A rapid diffusion of stereoscopic image acquisition devices is expected in the next years. Among the different potential applications that depth information can enable, in this paper we focus on its exploitation as a novel information source in the task of scene classification, and in particular to discriminate between indoor and outdoor images. This issue has not been addressed so far in the literature, probably because the extraction of depth information from two-dimensional images is a computationally demanding task. However, new-generation stereo cameras will allow a very fast computation of depth maps. We experimentally show that depth information alone provides a discriminant capability between indoor and outdoor images close to state-of-the art methods based on colour, edge and texture information, and that it allows to improve their performance, when it is used as an additional information source.

Unsupervised classemes

In this paper we present a new model of semantic features that, unlike previously presented methods, does not rely on the presence of a labeled training data base, as the creation of the feature extraction function is done in an unsupervised manner. We test these features on an unsupervised classification (clustering) task, and show that they outperform primitive (low-level) features, and that have performance comparable to that of supervised semantic features, which are much more expensive to determine relying on the presence of a labeled training set to train the feature extraction function.