|
|
3D Head Model Classification |
INVESTIGATORS
Dr. Hau San WONG, Dr. Bo MA, Ms. Xin TONG, Ms. Huiyang QU
KEYWORDS:
3D head model classification, Kernel Fuzzy c-means, Kernel Clustering-based Discriminant Analysis (KCDA)
EGI (Extended Gaussian Image), 2D subspace analysis, PCA, LDA
Evolutionary Strategy
FUNDING AGENCY
The Research Grants Council of Hong Kong Special Administrative Region, China
BRIEF DESCRIPTION
The 3D head model classification problem is addressed by use of several methods:
Optimized EGI
Our proposed method includes the following three steps: first, we compute the EGI as the original features for the 3D head models; second, we perform the 2D subspace analysis techniques to obtain optimized feature representation; finally, based on the optimized feature, we adopt a suitable classifier to perform the classification.
Fig. 1. Four example 3D head models in the database
Fig. 2. Results on the EGIs of size 20 by 30 using NMC Fig. 3. Results on EGIs of size 14 by 14 using NMC
Evolutionary Optimization of Feature Representation
We propose a new feature representation for 3D point cloud models based on a set of principal projection axes. The point set is then projected onto each of these axes, and a suitable summary statistics of the projected point set along each axis is calculated. The complete set of statistics is then adopted as the feature representation of the point set.
Based on this representation, we need to search for the optimal set of projection axes which can best distinguish the different classes of point cloud models in the database. In general, this optimization problem is difficult due to the size of the search space, in which we need to identify the optimal combination of projection directions among the vast number of different possibilities. As a result, we propose to adopt Evolutionary Strategy (ES) as the optimization technique, in view of the capability of ES to explore many regions of the search space in parallel through the generation of a large number of potential candidate solutions in a population. The effectiveness of each solution is measured through a fitness function which represents our current optimization criterion, and individuals associated with high fitness values are allowed to reproduce through the operations of recombination and mutation, while those with low fitness values are displaced out of the population[4]. In this way, highly effective solutions will gradually emerge through this process of competition and selection, so that we can get a satisfactory classification rate, while the efficiency and simplicity are also maintained.
Fig.1.Classification rate of the models Fig. 2. The classification comparison between
our ES-based method and four popular methods
Kernel Clustering-based Discriminant Analysis (KCDA)
This method works by first mapping the original data into another high-dimensional space, and then performing clustering-based discriminant analysis in the feature space. The main idea of clustering-based discriminant analysis is to overcome the Gaussian assumption limitation of the traditional linear discriminant analysis by using a new criterion that takes into account the multiple cluster structure possibly embedded within some classes. As a result, Kernel CDA tries to get through the limitations of both Gaussian assumption and linearity facing the traditional linear discriminant analysis simultaneously. Integrated with kernel fuzzy c-means, nearest mean classifier (NMC) and nearest prototype classifier (NPC), a group of tests of this method on 3D head model dataset have been carried out, reporting very promising experimental results.
Fig. 1. Examples in the 3D head model database.
1. Classification rates for 2-class problem using NMC.
20
30
40
50
60
LDA
0.8806
0.8794
0.9031
0.9133
0.9107
CDA
0.9056
0.9029
0.9375
0.9467
0.9536
KLDA
0.9111
0.9088
0.9344
0.9433
0.9464
KCDA
0.9472
0.9206
0.95
0.96
0.9679
Table 2. Classification rates for 4-class problem using NMC.
80
100
120
140
160
LDA
0.5438
0.63
0.6393
0.6846
0.6292
CDA
0.8844
0.8867
0.8831
0.8923
0.8958
KLDA
0.8813
0.8667
0.8786
0.9115
0.9125
KCDA
0.8969
0.8833
0.9034
0.9269
0.9208
Table 3. Classification rates for 8-class problem using NMC.
240
280
320
360
400
LDA
0.4339
0.4212
0.4
0.4114
0.4114
CDA
0.7964
0.7923
0.7896
0.7909
0.79
KLDA
0.85
0.8481
0.8542
0.8705
0.875
KCDA
0.8554
0.8558
0.875
0.875
0.8825
Table 4. Classification rates for 2-class problem using NPC.
20
30
40
50
60
LDA
0.8583
0.8794
0.9063
0.9233
0.9429
CDA
0.9278
0.9353
0.9594
0.96
0.9643
KLDA
0.9278
0.9206
0.9469
0.9567
0.9607
KCDA
0.9444
0.9412
0.9625
0.97
0.975
Table 5. Classification rates for 4-class problem using NPC.
80
100
120
140
160
LDA
0.7594
0.73
0.7679
0.7269
0.7458
CDA
0.9563
0.9467
0.9571
0.9615
0.9625
KLDA
0.95
0.9367
0.9536
0.95
0.9625
KCDA
0.9625
0.9667
0.9714
0.9654
0.9792
Table 6. Classification rates for 8-class problem using NPC.
240
280
320
360
400
LDA
0.3268
0.3432
0.3708
0.3636
0.3614
CDA
0.8304
0.8346
0.85
0.8614
0.8636
KLDA
0.8482
0.8519
0.8583
0.8705
0.875
KCDA
0.8589
0.8615
0.8688
0.8841
0.8925
PUBLICATIONS
Xin Tong, Hau-san Wong, Bo Ma, Horace H.S.Ip, "Evolutionary Optimization of Feature Representation for 3D Point-based Model Classification", Accepted by ICPR, HONG KONG,2006
Xin Tong, Hau-san Wong, Bo Ma, "3D Head Model Classification Using Optimized EGI", Proc. of SPIE-IS&T Electronic Imaging. SPIE Vol.6056.60560M, CA, USA, 2006
Bo Ma, Hui-yang Qu, Hau-san Wong, Yao Lu. 3D Head Model Classification Using KCDA. 2006 Pacific-Rim Conference on Multimedia (PCM 2006).
