Abstract
[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]


Correspondence between Human Shape Models using Curvature Flow


In this paper, we propose the method of obtaining the correspondence between the points on two human body models.

The 3D shape data of the objects, such as human bodies, have come to be easily obtained using 3D measuring devices such as range finders. In the apparel field, there is a demand of measuring human bodies automatically using their 3D shape data. It is needed to extract the measurement points, which are the reference points used in measurement, on human bodies automatically. The measurement points are, however, not always the geometrically characteristic points. It is difficult to extract the measurement points directly from the shape information of the human bodies.

In this paper, by using a standard human body model, the extraction of the measurement points is realized as following: (1)The standard model, which has the position information of the measurement points, is prepared; (2)The correspondence between the points on the standard model and those on the measured model generated from 3D shape data are obtained using the geometrically characteristic points; (3)The points on the measured model that correspond to the measurement points on the standard model are regarded as the measurement points of the measured model. The essential problem in this process is how to obtain the correspondence between the points on two human body models.

We assume that the pose of human body models is standing up straight, and the position and the direction are given. The human body models are represented as the patch models, and the correspondence between the points on the two human body models is represented by the correspondence between the vertices of the patch models.

For making correspondence between the points on the two human body models, the global features, which are common to the human body, have to be extracted. We use the human body surfaces' global unevenness for representing the features of human body shape, because it is common to almost all human bodies.

First, each model's surface is smoothed for removing the local unevenness and preserving the global one. It is desired that no unevenness region is generated and the global unevenness remains as the model is smoothed. In our method, each model's surface is smoothed by using the curvature flow. The degree of smoothing is called scale, and as the scale is larger, the model is smoother. Because the most appropriate scale depends on the regional unevenness, several models of different scales are generated.

Next, for each scale, the feature points are extracted from the model's vertices by calculating the curvature of the model surface. Giving priority to the feature points on the model of larger scale, the correspondence between the extracted feature points is found. The correspondence between the global feature points extracted from the model of the largest scale is found at first, and the correspondence between the feature points extracted from the model of smaller scale is found in sequence. A better result is expected by taking advantage of the correspondence in larger scales.

Finally, the correspondence about all the vertices is found out by using the obtained correspondence between the feature points. The vertices on the measured model that correspond to the measurement points on the standard model are obtained.

We conducted the experiment of applying our method to human shape models that are generated from the 3D shape data. We tried another smoothing method averaging the positions of neighboring vertices (simple smoothing). The smoothed results by the simple smoothing and the curvature flow smoothing were compared. As a result, the global unevenness was remaining in the case of the curvature flow smoothing, while new unevenness regions were generated in the case of the simple smoothing. In order to evaluate the obtained correspondence between the vertices, we performed the automatic measurement of human bodies by our method. And we compared its error with the error of the manual measurement by experts. As a result, the automatic measurement by our method had the close accuracy to the manual measurement by experts.


Go back to Thesis Page