Vector quantization based point cloud / image compression

An apparatus comprising:

• a processor;
• and
• a memory storing instructions operative, when executed by the processor, to cause the apparatus to:
• access a first bitstream,
• wherein the first bitstream represents a list of feature descriptors;
• obtain a codebook comprising one or more codewords,
• wherein the one or more codewords comprise a first set of one-hot indexes or vectors;
• obtain a probability distribution over the one or more codewords,
• wherein the probability distribution is learned by a neural network;
• decode a second set of one or more one-hot indexes from the first bitstream,
• wherein each one-hot index of the second set is associated with a feature descriptor in the list of feature descriptors;
• decode each feature descriptor based on the codebook and the respective one-hot index of the second set;
• and
• reconstruct an approximation of input data based on the decoded feature descriptors.

A method comprising:

• obtaining a list of feature descriptors;
• obtaining a codebook comprising one or more codewords,
• wherein the one or more codewords comprise a first set of one-hot indexes or vectors;
• determining a second set of one or more one-hot indexes by processing an extracted feature using a neural network-based encoder;
• obtaining a probability distribution over the one or more codewords,
• wherein the probability distribution is learned by a neural network;
• and
• encoding the second set of one or more one-hot indexes into a bitstream.