Supplementary Material: Unsupervised Pointcloud Registration via Differentiable Rendering

We only have one learned component in our model, the feature encoder, which is implemented using the ResNet basic block defined in the torchvision library. Our feature encoder takes as input an RGB image. The first layer is a 2D convolution layer with a kernel size of 3 and and output channel dimension of 64. This is followed by two ResNet basic blocks that retain the spatial and feature dimensions of the activations. Finally, we use a 2D convolution layer to map the feature dimension from 64 to 32. We reduce the feature dimension to 32 as it allows us to use the fast kNN CUDA kernel defined in PyTorch3D [1]. All convolution layers are followed by BatchNorm and ReLU activation, except for the last layer.