Vol.1 No.1 2008

45/76

Research paper : Technologies for the design and retail service of well-fitting eyeglass frames (M. Mochimaru, et al.)−42−Synthesiology - English edition Vol.1 No.1 (2008) proper size got significantly better scores for overall fit than the conventional frames (p<0.01). Moreover, the tightening force was significantly smaller in the new frames (p<0.01). On the other hand, no significant differences were observed in slip range. Thus, the new frames fit without slip and had a smaller tightening force. 3.3 3-D head shape reconstruction from multiple images[5]Location in the subject distribution map could be estimated from 15 face dimensions, and a suitable size of eyeglass frame could be selected. An expert anthropometrist is necessary to obtain reliable face dimensions, thus it is difficult to use this method in actual retail shops. In addition, 15 face dimensions are not sufficient for reconstructing 3-D face shape, which is used for computer graphic representation (Figure 2), while the special 3-D face scanner mentioned in Section 3.1 is too large and expensive to use in retail shops. Therefore, a novel measurement method for human body shape was developed.The variation in human bodies can be represented in the eigenspace with a smaller number of independent dimensions (Section 3.1). Namely, any human body shape can be described using only dozens of principal components rather than millions of data points. Therefore, we developed a new method for reconstructing, rather than measuring, the 3-D head shape of an individual. The average head shape is used as the initial shape, and unknown values (principal component scores describing the individual shape) were obtained by minimizing the differences between camera images and the projected image of the 3-D head shape reconstructed from principal component scores (A(n) in equation (1)) Reconstructed 3-D head shape was projected into image planes based on calibration parameters of multiple cameras. The difference between the real 2-D image and the projected image was evaluated by the edge distance of two images. The principal component scores A(n) were optimized to minimize image differences, and the optimal 3-D shape was obtained from the optimized principal component scores (Figure 7). The calculated 3-D head shape was compared with 3-D shape measured by the head scanner (Section 3.1), and the average error was found to be 2.0 mm.This technology is not a measurement technology, but is a reconstruction technology using multiple cameras based on the database of homologous body shapes. However, this technology has a similar function as a measurement technology, because 3-D head shape can be obtained from multiple cameras to an accuracy of 2.0 mm. As this method is not based on triangulation between a camera and projector, patterned light projection and scanning processes are unnecessary. A homologous model of an individual can be obtained by simultaneously capturing images with multiple cameras. Face dimensions are extracted from the model, and used for recommendation of eyeglass frame size. Furthermore, even when part of the head is occluded, the reconstructed model is always complete.3.4 Computational estimation of the impression for the face and frame[6,7]Size recommendation can be realized by size variation design technology (Section 3.2) and individual head shape measurement technology (Section 3.3). In this section, style recommendation technology is described. This technology is for estimating the impression ratings of the combination of an individual face and a selected eyeglass frame.Words were used to describe the impression of the face and the eyeglass frame. In a preliminary experiment, we selected 42 adjectives from about 100 words obtained from a dictionary. Photographs of 10 different faces with different eyeglasses were presented to young adult female subjects and impression of each photograph was rated against the 42 adjectives. The ratings were analyzed using a factor analysis, and 4 factors were extracted. The 4 factors were interpreted as “gentle - stern”, “cool - hot”, “cheerful - quiet” and “younger looking - older looking”. The following 3 pairs of adjectives were added based on the original marketing research of the collaborating eyeglass frame maker: “sophisticated - unstylish”, “natural - unnatural” and “optimistic - nervous”. Impression ratings for these 7 pairs of words were obtained for images of a face and frame by experiments using sensory evaluation. We hypothesized that impression rating could be estimated as a function of shape factors for the face and eyeglass frame. Therefore, we identified this function based on experimental data, and validated the performance of this function using actual data.Generated virtual face images were used for the experiment. Virtual face images were generated to cover the variations in young adult male Japanese faces using the database of Fig. 7 Reconstruction of 3-D face and head shape from multiple images

※このページを正しく表示するにはFlashPlayer9以上が必要です