Hardware-Accelerated Silhouette Matching
Hardware-Accelerated Silhouette Matching
复制标题
硬件加速轮廓匹配
DOI：
--
复制
发表时间：
2000
期刊：
影响因子：
--
通讯作者：
H. Seidel
中科院分区：
文献类型：
--
作者：
H. Lensch;W. Heidrich;H. Seidel
提出了一种硬件加速算法，用于计算图像中的轮廓与投影3D模型的轮廓之间的误差，从而估计摄像机的位置。由于渲染算法在过去几年中变得更加复杂，对逼真模型的需求也增加了，特别是对于现实世界对象的模型。因此，不仅3D形状而且表面颜色或反射率也是主要感兴趣的。通常通过产生例如三角形网格的扫描设备来获得3D模型。在使用相机的第二过程中经常收集颜色信息。然后，生成整个模型需要组合3D形状和2D图像/纹理。必须为每个图像计算相机相对于3D网格的位置和方向，以便将纹理缝合到3D曲面上。在以前的工作[1，2]中，摄像机设置是通过最小化2D图像中对象的轮廓与投影3D模型的轮廓之间的误差来确定的。不同于以前的算法，误差是通过像素到轮廓距离的总和来估计的，我们的方法直接测量被两个轮廓包围的面积。由于3D模型的投影可以使用图形硬件来计算，因此获得这种面积测量是非常容易和有效的：作为预处理步骤，2D图像被分割成属于背景的像素(黑色)和被对象覆盖的像素(白色)。首先，对于每个相机位置，3D模型在黑色背景前被渲染为白色，并且使用每像素异或运算将结果与分割图像组合。现在，轮廓之间的像素准确地保持为白色，它们的数量可以通过简单地计算直方图来计算。由于所有这些步骤都可以在硬件中执行，因此显着加快了相机参数的非线性优化。此外，在不同的图像分辨率上进行分层优化，以减少渲染工作量。给出相机的焦距和物体距离的粗略估计，甚至可以在几分钟内获得相机的位置和旋转，而无需任何用户交互。
A hardware-accelerated algorithm is presented to compute the error between a silhouette in an image and the silhouette of a projected 3D model to estimate the camera position. Since rendering algorithms have become more sophisticated in the last years also the demand for realistic models has been increased, especially for models of real world objects. Hereby, not only the 3D shape but also the surface color or reflectance is of major interest. The 3D model is commonly obtained by a scanning device producing for example a triangle mesh. Color information is frequently collected during a second process using a camera. Generating the entire model then requires to combine the 3D shape and the 2D images/textures. The camera position and orientation relative to the 3D mesh must be computed for each image in order to stitch the textures onto the 3D surface. In previous work [1, 2] the camera settings are determined by minimizing the error between the outline of the object in the 2D image and the outline of the projected 3D model. In contrast to previous algorithms where the error is estimated by a sum of pixel to outline distances, our approach directly measures the area enclosed by the two outlines. Since the projection of the 3D model can be computed using graphics hardware it is very easy and efficient to obtain this area measurement: as a preprocessing step the 2D image is segmented into pixels belonging to the background (black) and those covered by the object (white). For each camera position at first the 3D model is rendered colored white in front of a black background and the result is combined with the segmented image using a per pixel XOR-operation. Now, exactly the pixel between the outlines remain white and their number can be counted by simply evaluating a histogram. Since all these steps can be executed in hardware the non-linear optimization for the camera parameters is significantly accelerated. Further, the optimization is done hierarchically on different image resolutions to decrease the rendering effort. Given the focal length of the camera and a rough estimate for the object distance it is even possible to obtain the camera position and rotation without any user interaction in a few minutes.