Network, System and Method for 3D Shape Generation

A network for generating 3D shape, including

• an image feature network containing successively five convolutional layer groups, and four pooling layers sandwiched by the five convolutional layer groups;
• an initial ellipsoid mesh, and
• a cascaded mesh deformation network containing successively three deformation blocks, and two graph unpooling layers sandwiched by the three successive deformation blocks;
• wherein the first convolutional layer group contains two successive convolutional layers each having a 3×3 the convolutional kernel size and 64 channels;
• the second convolutional layer group contains two successive convolutional layers each having a 3×3 the convolutional kernel size and 128 channels;
• the third convolutional layer group contains three successive convolutional layers each having a 3×3 the convolutional kernel size and 256 channels;
• the fourth convolutional layer group contains three successive convolutional layers each having a 3×3 the convolutional kernel size and 512 channels;
• and the fifth convolutional layer group also contains three successive convolutional layers each having a 3×3 the convolutional kernel size and 512 channels;
• the four pooling layers each has a 2×2 pooling core size;
• the initial ellipsoid mesh is centered at 0.8 m in front of a camera, and the radius of its three axis is respectively 0.2 m, 0.2 m, and 0.4 m;
• wherein the camera represents a coordinates system in which its original position is the view point;
• the deformation blocks each comprises
• a perceptual feature pooling layer,
• a graph based ResNet (G-ResNet) consisting of 14 graph residual convolutional layers with 128 channels, and
• a loss module calculating an overall loss (lall) that is a weighted sum of the Chamfer loss (lc), the normal loss (ln), the Laplacian regularization (llap), and the edge length regularization (lloc) in accordance the formula lall=lc+1.6e−4*ln+0.3*llap+0.1*lloc;
• and
• the graph unpooling layer each is an edge-based unpooling layer in which a vertex is added at the center of each edge of the input mesh and the 3D feature of the newly added vertex is set as the average of the two end points of the edge, so that the three vertices added on a same original triangle are connected to create four new triangles from the original triangle.