Human Face Recognition
|Advantages and disadvantages of 3D face recognition|
Advantages and disadvantages of 3D face recognition
As previously discussed, face recognition using 2D images is sensitive to illumination changes. The light collected from a face is a function of the geometry of the face, the albedo of the face, the properties of the light source and the properties of the camera. Given this complexity, it is difficult to develop models that take all these variations into account. Training using different illumination scenarios as well as illumination normalization of 2D images has been used, but with limited success. In 3D images, variations in illumination only affect the texture of the face, yet the captured facial shape remains intact
Another differentiating factor between 2D and 3D face recognition is the effect of pose variation. In 2D images effort has been put into transforming an image into a canonical position. However, this relies on accurate landmark placement and does not tackle the issue of occlusion. Moreover, in 2D this task is nearly impossible use to the projective nature of 2D images. To circumvent this problem it is possible to ore different views of the face. This, however, requires a large number of 2D images from many different views to be collected. An alternative approach to address the pose variation problem in 2D images is either based on statistical models for view interpolation or on the use of generative models. Other strategies including sampling the plenoptic function of a face using lightfield techniques. Using 3D images, this view interpolation can be simply solved by re-rendering the 3D face data with a new pose. This allows a 3Dmorphable model to estimate the 3D shape of unseen faces from non-frontal 2D input images and to generate 2D frontal views of the reconstructed faces by re-rendering. Another pose-related problem is that the physical dimensions of the face in 2D images are unknown. The size of a face in 2D images is essentially a function of the distance of the subject from the sensor. However, in 3D images the physical dimensions of the face are known and are inherently encoded in the data. In contrast to 2D images, 3D images are better at capturing the surface geometry of the face. Traditional 2D image-based face recognition focuses on high-contrast areas of the face such as eyes, mouth, nose and face boundary because low contrast areas such as the jaw and cheeks are difficult to describe from intensity images. 3D images, on the other hand, make no distinction between high- and low-contrast areas. 3D face recognition, however, is not without its problems. Illumination, for example, may not be an issue during the processing of 3D data, but it is still a problem during capturing. Depending on the sensor technology used, oily parts of the face with high reflectance may introduce artifacts under certain lighting on the surface. The overall quality of 3D image data collected using a range camera is perhaps not as reliable as 2D image data, because 3D sensor technology is currently not as mature as 2D sensors. Another disadvantage of 3D face recognition techniques is the cost of the hardware. 3D capturing equipment is getting cheaper and more widely available but its price is significantly higher compared to a high resolution digital camera. Moreover, the current computational cost of processing 3D data is higher than for 2D data.
Finally, one of the most important disadvantages of 3D face recognition is the fact that 3D capturing technology requires cooperation from a subject. As mentioned above, lens or laserbased scanners require the subject to be at a certain distance from the sensor. Furthermore, a laser scanner requires a few seconds of complete immobility, while a traditional camera can capture images from far away with no cooperation from the subjects. In addition, there are currently very few high-quality 3D face databases available for testing and evaluation purposes. Those databases that are available are of very small size compared to 2D face databases used for benchmarking.
The comparison of different 3D face recognition techniques is very challenging for a number of reasons: Firstly, there are very few standardized 3D face databases which are used for benchmarking purposes. Thus, the size and type of 3D face datasets varies significantly across different publications. Secondly, there are differences in the experimental setup and in the metrics which are used to evaluate the performance of face recognition techniques. Table 3.4 gives an overview of the different methods discussed in the previous section, in terms of the data and algorithms used and the reported recognition performance. Even though 3D face recognition is still a new and emerging area, there is a need to compare the strength of each technique in a controlled setting where they would be subjected to the same evaluation protocol on a large dataset. This need for objective evaluation prompted the design of the FRVT 2000 and FRVT 2002 evaluation studies aswell as the upcoming FRVT 2006 (http://www.frvt.org/). Both studies follow the principles of biometric evaluation laid down in the FERET evaluation strategy (Phillips et al., 2000). So far, these evaluation studies are limited to 2D face recognition techniques but will hopefully include 3D face recognition techniques in the near future.
Want To Know more with
Contact for more learning: webmaster@freehost7com
The contents of this webpage are copyrighted © 2008 www.freehost7.com
All Rights Reserved.