Supplementary : 3 DN : 3 D Deformation Network

We use an ADAM optimizer with initial learning rate 0.0005, momentum 0.9 and batch size 4. Our network is implemented with Tensorflow and trained on an Nvidia GTX1080 Ti GPU. When the input is a point cloud, we have used a point cloud of size 2048 × 3. No Batch Normalization layer is used. The weights for the losses are ωL1 = 1000, ωL2 = 1, ωL3 = 10000, ωL4 = 1, ωL5 = 1000, ωL6 = 0.01, ωL7 = 1000. The mesh sampling operator is implemented with CuDA acceleration and Tensorflow. Since number of vertices are varying for different meshes and we need to train the network with batch size > 1, we set a maximum number of vertices and triangles to make the operation trainable with batch size > 1, and we also input the number of vertices and triangles for each sample. In the CuDA implementation, we sample the same number of points for meshes with different number of vertices. The inputs to the mesh sampling operator are V ∈ RB×NV max×3, T ∈ ZB×NTmax×3, NV ≤ NVmax ∈ ZB×1, NT ≤ NTmax ∈ ZB×1, wV 1 ∈ (0, 1)B×NTmax×3, wV 2 ∈ (0, 1)B×NTmax×3, wV 3 ∈ (0, 1)B×NTmax×3, where B is batch size, NVmax is the max number of mesh vertices,, NTmax is the max number of mesh faces,, NV and NT indicate the number of vertices and triangles for each sample in the mini batch, wV 1, wV 2 and wV 3 are the random barycentric coordinate weights.