Dynamic time warping for 3d mesh sequences retrieval

Postgraduate Thesis uoadl:1317039 602 Read counter

Unit:
Κατεύθυνση / ειδίκευση Επεξεργασία-Μάθηση Σήματος και Πληροφορίας (ΕΜΠ)
Library of the School of Science
Deposit date:
2013-09-02
Year:
2013
Author:
Βεϊνίδης Χρήστος
Supervisors info:
Θεοχάρης Θεοχάρης Καθηγητής ΕΚΠΑ, Ιωάννης Πρατικάκης Επίκ. Καθηγητής Δημοκρίτειου Πανεπιστήμιου Θράκης
Original Title:
Δυναμική στρέβλωση χρόνου για την ανάκτηση 3δ πλεγματοσειρών
Languages:
Greek
Translated title:
Dynamic time warping for 3d mesh sequences retrieval
Summary:
A 3D mesh sequence consists of a number of mesh objects in order. The retrieval
of 3D mesh sequences which represent human actions is attracting increasing
research interest. The process of 3D mesh object retrieval generallycan be
separated into three parts; the pose normalization of the object in order to
remain invariant to probable scale, rotation and translation transforms, the
extraction of an object’s descriptor and the evaluation of the similarity
between the particular object and a list of other objects. The descriptor is a
short but meaningful representation of the mesh object and it is possible that
the object’s descriptor is independent from the above transformations, so the
first part of the process (the normalization part) maybe optional.The third
part (the evaluation of the similarities) is as important as the other parts,
because the choice of the distance metric that is used for the computation of
the similarities between the descriptors of the models, is crucial.
Particularly, in the case of mesh sequences, new methods for the evaluation of
the distance must be sought. The most considerable section of this work is
related to the third part of the retrieval process. Specifically, the object
descriptor that was usedwas implemented in the past and incorporates pose
normalization methods. This descriptor is called hybrid and was created in
order to be applied in static mesh objects, so it was necessary to seek ways to
extend it and make it appropriate for the case of the mesh sequences. So, for
each of the mesh objects which constitute the mesh sequences, the hybrid
descriptor was extracted and two different methods were adopted for the purpose
of computing the distance between two sequences. The first relies on an
exhaustive search of the maximum similarity between the objects of two
sequences and presupposes the computation of the L1 distance between all the
object descriptors from the first sequence andall the object descriptors from
the second. The second method relies on the Dynamic Time Warping method
andregards a mesh sequence as a curve which is defined in a space whose
dimensionality is determined by the number of the coordinates of the vectors
extracted using the hybrid descriptor. Experimental evaluation showed that the
percentage of successful retrievals is much higher using the second method, for
all the metrics used to evaluate the performance of the methods.
Keywords:
Dynamic Time Warping, 3D mesh sequences, Retrieval, Descriptor, Distance metric
Index:
Yes
Number of index pages:
9-11
Contains images:
Yes
Number of references:
44
Number of pages:
66
File:
File access is restricted.

document.pdf
2 MB
File access is restricted.