Indexation 3D

A Novel Efficient 3D Object Retrieval  Method

Based on Representative Slices

Ilyass OUAZZANI TAYBI 1,∗, Rachid ALAOUI 2, Fatima RAFII ZAKANI 1, Khadija ARHID 1, Mohcine BOUKSIM 1  and Taoufiq  GADI 1

1: Laboratory Informatics,  Imaging and Modeling of Complex Systems (LIIMSC) Faculty of Sciences and Techniques,  Hassan 1st University Settat, Morocco

2: Laboratory of Systems Engineering and Information Technology  (LISTI) ENSA, Ibn Zohr University Agadir, Morocco

Email : ilyass.ouazzani@gmail.com

Abstract—In  the  last  few  years,  the  request  for  a  content- based  3D object  retrieval system has become a significant issue. At this  point,  the  principal challenge  is the  mapping  of the  3D objects  into  compact  representations referred to as descriptors, which  serve  as  search  keys  over  the  retrieval process.  In  this paper,  a new approach will be proposed  for 3D objects indexing and  retrieval.  The  main  idea  is  to  normalize   the  3D  objects to insure  invariance with respect to affine transformations, and then  characterize them  by  a  set  of  representative  slices  (RS) along their three principal axes, transforming the shape-matching problem  between  3D objects  into  similarity  measuring  between their  representative slices. In order  to reduce the time required to  search  without  diminishing  the  relevance of the  results, we choose among  the  extracted  slices from  the  3D object  the  ones that  give the  best  representation. To  achieve  this  task,  we use the  k-means  clustering   method  to  pull  out  the  representative slices. For the presentation of the effectiveness and superiority of our approach we conduct  a comparison  of our approach against

3D Zernike  descriptor on 146 3D objects  from  Princeton Shape Benchmark (PSB) database.   Experiment results show that  our proposed  method  is superior  to 3D Zernike  descriptor.

Index Terms—K-means  clustering  method, 3D object indexing,

3D object  retrieval, 2D slices.

I.  INTRODUCTION

As a result of the increasing popularization of the Internet, together with the rapid development of 3D scanning technolo- gies and modeling tools, content-based 3D object retrieval has become an active research field that has attracted a significant amount  of interest.  Generally,  the ultimate  aim of content- based  3D model  retrieval  systems  is to approximate  human visual perception so that semantically  similar 3D models can be correctly  retrieved based on their looks.

Indeed, the content-based  means that the retriever uses the visual features of 3D objects themselves, rather than relying on human-inputted metadata such as captions or keywords. In reality, the human-inputted metadata, as simple as it seems at first glance, has many drawbacks,  including the time needed for the labeling  of the collection, and the subjective  aspect of the keywords used, these keywords are necessarily  related to collection and culture of  the operator  carrying out the labeling. Therefore, the visual features of 3D objects should be automatically or semi-automatically  extracted and expected to characterize their contents.

The essential processing flow of a content-based 3D object retrieval  system can be roughly  described  as follows: the compact  and representative  features  are first computed  and extracted  automatically  from 3D objects to build  their mul- tidimensional indices. The similarity or dissimilarity measure between a query and each target object  in the database is then defined and calculated  in the multidimensional feature space. The similarity  values are then sorted in descending  order so that the models having the largest similarity values are returned as the matching results, on the basis of which browsing and retrieval in 3D object databases are finally implemented.

In this paper we present a new approach to describe a 3D object based on representative slices. The proposed approach consists of five steps. First, we normalize 3D objects to ensure translation, scale and rotation invariance  of descriptors. Next, we construct the initial set of 2D slices by taking, for each

3D object, slices following  the three principal  axes. Then, for each slice, we use Hus invariant  moments  to compute numerical signature. Thereafter, we select slices that give the best representation  of the 3D object by using the k-means clustering method. Finally, we compute the similarity between representative slices of query and representative  slices of each target object in the database using Hausdorff distance.

The  outline of  the rest of  this paper is  structured  as follow: in the next section, we present a state of the art by classifying existing indexing methods. In section 3, we discuss the proposed  approach.  Section 4  shows the experimental results in detail for 3D objects’ retrievals. Finally, section 5 concludes the proposed 3D object’ approach and recommends some future works.

II.  STATE OF THE ART

In this section, we provide an overview of the related work in the field of 3D shape descriptors for 3D object retrieval. In this context, we classify existing approaches into four groups:

-  The methods based on the information conveyed by the

3D object geometry;

-  The methods  using 2D  projections  of  the 3D  object

associated  with the information of depth;

-  The methods that ultimately require information  in two

dimensions;

978-1-5090-5146-5/16/$31.00 ©2016 IEEE

-  The methods combining several descriptors to character- ize a 3D object.

We invite the reader  to consult  the work of [1] [2] [3], which provide a comparison between different approaches of indexing and retrieval of 3D objects.

A. 3D / 3D approaches

3D based methods for 3D object retrieval involve all meth- ods that take into consideration  the 3D model as itself to retrieve  information  and define the descriptor. The choice of this signature shows five groups within the 3D approaches.

The global-based  methods  present  approaches  where the descriptor characterizes the whole 3D object. Osada et al. [4] [5] represent the signature of an object as a shape distribution sampled  from a shape function  measuring  global geometric properties of the object. The shape descriptor of a 3D object is given by a probability distribution that counts the occurrence of  Euclidean distances between pairs of  points randomly chosen on the surface of the object.

...