Faster, Better and More Detailed: 3D Face Reconstruction with Graph Convolutional Networks.

For the 2D image encoder in coarse mesh reconstruction network, we manage to adopt both ResNet-50 [1] and MobileNetV2 [2] for our purpose, with almost the same parameter settings and training procedure. We illustrate our system again in Fig. 1 for ease of reference. We denote spiral convolution layer with h hops, t filters and v vertices as SPConv(h,t,v), pooling and unpooling by a factor of p as Unpool(p) and Pool(p) respectively. FC layer is written as FC(d), where d is the output feature dimension. It has be mentioned that, in Fig. 1 as well as the following paragraphs, we take ResNet-50 as the example. Our spiral mesh decoder takes a 256-d feature embedding from image encoder as input, and has the following structure: FC(52*128)→Unpool(4)→SPConv(1,128,208)→Unpool(4)→SPConv(1,64,832) →Unpool(4)→SPConv(2,32,3326)→Unpool(4)→SPConv(2,16,13304) →Unpool(4)→SPConv(2,8,53215)→SPConv(2,3,53215). For simplicity, we describe the image feature sampling, adaptation and summation as one operation, viz. AddF(r,d,v), where r represents the spatial resolution of image features. Our refinement network uses a concatenation of coarse face normals and per vertex RGB values as input, and can be described as: SPConv(2,8,53215)→Pool(4)→SPConv(2,16,13304)→Pool(4) →SPConv(2,32,3326)→AddF(64,32,3326)→Unpool(4)→SPConv(2,16,13304) →AddF(128,16,13304)→Unpool(4)→SPConv(2,8,53215)→SPConv(2,3,53215). ELU [3] activation is used after each spiral convolution and fully connected layer, except for the last layer before getting the coarse mesh or per vertex shape displacement prediction.