Point cloud attribute compression with neural implicit representations

Point cloud attribute compression is a challenging task due to the irregularity of the point cloud domain. This makes it difficult to extend traditional compression principles like transform coding to such data type, often requiring complex and sophisticated schemes. At the same time, deep learning is gaining popularity as a way to design optimized compression algorithms. Existing end-to-end signal compression schemes using neural networks are largely based on an autoencoder-like structure, where a universal encoding function creates a compact latent space and the signal representation in this space is quantized and stored. In this paper, we follow a different approach by adopting neural implicit representation networks, i.e., neural networks that are queried with a coordinate and returns the signal value at that coordinate. A network of this kind is trained to overfit the signal to be compressed and the neural network itself, in its weights and biases, becomes the compressed representation of the signal. Efficient techniques to quantize neural network weights are then used to limit the rate of the compressed representation. We also show that it is possible to induce prior knowledge about the class of signals of interest via meta-learning techniques, thus providing an initialization value for the network weights. This procedure has a twofold advantage in terms of complexity and compression efficiency. In particular, it allows to finetune the network for the representation of the specific signal of interest with a small number of iterations, limiting encoding complexity. Moreover, the weights can be encoded differentially with respect to such initialization to achieve greater rate-distortion efficiency. Preliminary experiments show that the proposed method is competitive with the latest G-PCC MPEG standard for point cloud attribute compression, and outperforms RAHT, a recent state-of-the-art method.