高斯导数-脸部朝向.pdf

Estimating Face orientation from Robust Detection of Salient Facial Structures Nicolas Gourier Daniela Hall James L. Crowley Prima Project, GRAVIR Laboratory, INRIA Rhone-Alpes,ˆ Montbonnot, FRANCE Abstract pal components analysis (PCA) [12]. However, projection highly sensitive to small changes in face position and im- Thispaper addresses the pr oblem of estimatingface ori- age scale, as well as partial occlusions and as a result has entation fr om automatic detection of salient facial struc- proved unusable in real systems. In general, global tech- tures using learned robustfeatur es. Face imagettes are de- niques such as projection to a principle components space tected using color and described using a weighted sum of tend to be sensitive to partial occlusions as well as changes locally normalized Gaussian receptive fields. Robust face in identity. featur es are learned by clustering the Gaussian derivative An alternative is to measure the relative position of responses within a training set face imagettes. The most salient anatomical facial structures such as eyes and lips [2]. reliable clusters are identified and used as featur es for de- The challenge is that such facial structures are difficult to tecting salient facial structures. We have found that a sin- detect in a general manner. Most authors rely on complex gle cluster is sufficient to pr ovide a detector for salient fa- adhoc operations that tend to be highly sensitive to environ- cial structures that is robust to face orientation, illumina- mental conditions. tion and identity. We describe how clusters are learned and We define salient featur