computer animation Y.GUO ET AL virtual worlds 00●年◆。eee。ee。ee00。。●。。。.e.e.................0g Figure 8.Some models that cannot be detected. This verifies our subjective perception,that is,geometric features of primary facial organs are the most distinct characteristics that 3D faces differ from other models. We further test the detection accuracy of the learned classifiers on a probe set,which includes 5323D faces and Figure 6.The first six most discriminative features selected by 1000 non-faces.Among them,half of the 3D faces still are synthesized using USF HumanID 3D face database with AdaBoost. different parameters;some are scanned using a 3Space Fast Scan hand held scanner,while all the rest models and fit into a common frame.Thereafter,their surfaces come from the public NTU model database.22 are voxelized in the frame with uniform voxel size and The detection time for a geometric model normally consistent orientation.Finally,geometric contexts with takes less than 10 seconds on a standard PC with an a set of variant edge length R over surface voxels are Intel PIV 3.0 GHz CPU and 1G memory.Figure 7.shows calculated for face detection using Adaboost learning. the receiver operator characteristic (ROC)curve of our detector.In addition,nearly 20 3D face or human models Face Detection in NTU database can not be detected due to influence of cap,glasses,or lack of features on 3D faces(Figure 8) The Adaboost learning algorithm described in Sec- tion Learning-Based 3D Face Model Detection'is applied to the training data.During this process,162 most Conclusions and Future discriminative classifiers are selected.Figure 6 shows Six Work features of them.As can be seen these features encode the primary facial features,for example,nose,canthus,etc. This paper presents an efficient 3D face detection approach based on AdaBoost learning.To compactly ROC curve 1.0 encode the facial geometric features,we introduce geometric context,a novel 3D shape descriptor,whose computation speed over model surface can be very fast 0.9 with integral volume.AdaBoost is employed to select the most discriminative geometric context-based features, 0.8 and to integrate them into a strong classifier for 3D face and non-face classification. 0.7 Generally,the detection rate of learning-based algorithm depends heavily on the training examples,for example,the number of examples as well as different 0.6 race.We intend to collect more 3D face models to enrich the training examples and to enhance the detection 0.5 performance. 0 10 20 30 40 50 False positives Our 3D face detection approach can be behaved as a preliminary step for 3D face recognition and biometric Figure 7.Roc curve for our detector on test set. identification,and can also be extended to the automatic ........... Copyright 2007 John Wiley Sons,Ltd. 490 Comp.Anim.Virtual Worlds 2007;18:483-492 DOL:10.1002/cavY. GUO ET AL. ........................................................................................... Figure 6. The first six most discriminative features selected by AdaBoost. and fit into a common frame. Thereafter, their surfaces are voxelized in the frame with uniform voxel size and consistent orientation. Finally, geometric contexts with a set of variant edge length R over surface voxels are calculated for face detection using Adaboost learning. Face Detection The Adaboost learning algorithm described in Sec￾tion ‘Learning-Based 3D Face Model Detection’ is applied to the training data. During this process, 162most discriminative classifiers are selected. Figure 6 shows Six features of them. As can be seen these features encode the primary facial features, for example, nose, canthus, etc. Figure 7. Roc curve for our detector on test set. Figure 8. Some models that cannot be detected. This verifies our subjective perception, that is, geometric features of primary facial organs are the most distinct characteristics that 3D faces differ from other models. We further test the detection accuracy of the learned classifiers on a probe set, which includes 532 3D faces and 1000 non-faces. Among them, half of the 3D faces still are synthesized using USF HumanID 3D face database with different parameters; some are scanned using a 3Space Fast Scan hand held scanner, while all the rest models come from the public NTU model database.22 The detection time for a geometric model normally takes less than 10 seconds on a standard PC with an Intel PIV 3.0 GHz CPU and 1G memory. Figure 7. shows the receiver operator characteristic (ROC) curve of our detector. In addition, nearly 20 3D face or human models in NTU database can not be detected due to influence of cap, glasses, or lack of features on 3D faces (Figure 8) Conclusions and Future Work This paper presents an efficient 3D face detection approach based on AdaBoost learning. To compactly encode the facial geometric features, we introduce geometric context, a novel 3D shape descriptor, whose computation speed over model surface can be very fast with integral volume. AdaBoost is employed to select the most discriminative geometric context-based features, and to integrate them into a strong classifier for 3D face and non-face classification. Generally, the detection rate of learning-based algorithm depends heavily on the training examples, for example, the number of examples as well as different race. We intend to collect more 3D face models to enrich the training examples and to enhance the detection performance. Our 3D face detection approach can be behaved as a preliminary step for 3D face recognition and biometric identification, and can also be extended to the automatic ............................................................................................ Copyright © 2007 John Wiley & Sons, Ltd. 490 Comp. Anim. Virtual Worlds 2007; 18: 483–492 DOI: 10.1002/cav