IATMLRP 2012 Abstracts


Full Papers
Paper Nr: 5
Title:

BASS TRACK SELECTION IN MIDI FILES AND MULTIMODAL IMPLICATIONS TO MELODY

Authors:

Octavio Vicente and Jose M. Iñesta

Abstract: Standard MIDI files consist of a number of tracks containing information that can be considered as a symbolic representation of music. Usually each track represents an instrument or voice in a music piece. The goal for this work is to identify the track that contains the bass line. This information is very relevant for a number of tasks like rhythm analysis or harmonic segmentation, among others. It is not easy since a bass line can be performed by very different kinds of instruments. We have approached this problem by using statistical features from the symbolic representation of music and a random forest classifier. The first experiment was to classify a track as bass or non-bass. Then we have tried to select the correct bass track in a multi-track MIDI file. Eventually, we have studied the issue of how different sources of information can help in this latter task. In particular, we have analyzed the interactions between bass and melody information. Yielded results were very accurate and melody track identification was significantly improved when using this kind of multimodal help.

Paper Nr: 10
Title:

MULTIPLE OBJECT TRACKING WITH RELATIONS

Authors:

Luca Cattelani, Cristina Manfredotti and Enza Messina

Abstract: Dealing with multi-object tracking raises several issues; an essential point is to model possible interactions between objects. Indeed, while reliable algorithms for tracking multiple non-interacting objects in constrained scenarios exist, tracking of multiple interacting objects in uncontrolled scenarios is still a challenge. The multiple-object tracking problem can be broken down into two subtasks: the detection of target objects, and the association between objects along time. Interaction between objects can yield erroneous associations that cause the interchange of object identities, therefore, the explicit recognition of the relationships between interacting objects in the scene can be useful to better detect the targets and understand their dynamics, making tracking more accurate. To make inference in relational domains we have developed an extension of particle filter, called relational particle filter, able to track simultaneously the objects in the domain and the evolution of their relationships. Experimental results show that our method can follow the targets’ path more closely than standard methods, being able to better predict their behaviours while decreasing the complexity of the tracking.

Paper Nr: 12
Title:

BAYESIAN SUPERVISED IMAGE CLASSIFICATION BASED ON A PAIRWISE COMPARISON METHOD

Authors:

F. Calle-Alonso, J. P. Arias-Nicolás, C. J. Pérez and J. Martín

Abstract: In this work, a novel classification method is proposed. The method uses a Bayesian regression model in a pairwise comparison framework. As a result, we obtain an automatic classification tool that allows new cases to be classified without the interaction of the user. The differences with other classification methods, are the two innovative relevance feedback tools for an iterative classification process. The first one is the information obtained from user after validating the results of the automatic classification. The second difference is the continuous adaptive distribution of the model’s parameters. It also has the advantage that can be used with problems with both a large number of characteristics and few number of elements. The method could be specially helpful for those professionals who have to make a decision based on images classification, such as doctors to determine the diagnosis of patients, meteorologists, traffic police to detect license plate, etc.

Paper Nr: 13
Title:

RESTRUCTURING VERSUS NON RESTRUCTURING INSERTIONS IN MDF INDEXES

Authors:

Aureo Serrano, Luisa Micó and Jose Oncina

Abstract: MDF tree is a data structure (index) that is used to speed up similarity searches in huge databases. To achieve its goal the indexes should exploit some property of the dissimilarity measure. MDF indexes assume that the dissimilarity measure can be viewed as a distance in a metric space. Moreover, in this framework is assumed that the distance is computationally very expensive and then, counting distance computations is a good measure of the time complexity. To tackle with a changing world, a problem arises when new points should be inserted in the index. Efficient algorithms should choose between trying to be efficient in search maintaining the “ideal” structure of the index or trying to be efficient when inserting but worsening the search time. In this work we propose an insertion algorithm for MDF trees that focus on optimizing insertion times. The worst case time complexity of the algorithm only depends on the depth of the MDF tree. We compare this algorithm with a similar one that focuses on search time performance. We also study the range of applicability of each one.

Paper Nr: 15
Title:

TWO METHODS FOR FILLED-IN DOCUMENT IMAGE IDENTIFICATION USING LOCAL FEATURES

Authors:

Diego Carrion-Robles, Vicent Castello-Fos, Juan-Carlos Perez-Cortes and Joaquim Arlandis

Abstract: In this work, the task of document image classification is dealt with, particularly in the case of pre-printed forms, where a large part of the document can be filled-in with the result of a potentially very different image. A method for the selection of discriminative local features is presented and tested along with two different classification algorithms. The first one is an incremental version of the method proposed in (Arlandis et al., 2009), based on similarity searching around a set anchor points, and the second one is based on a direct voting scheme ((Arlandis et al., 2011)). Experiments on a document database consisting of real office documents with a very high variability, as well as on the NIST SD6 database, are presented. A confidence measure intended to reject unknown documents (those that have not been indexed in advance as a given document class) is also proposed and tested.