Publication
The use of immersive representations for digital media has been gaining growing attention fueled by novel applications such as virtual and augmented reality. Point clouds, which represent shapes as a collection of unconnected points in space with geometric coordinates and associated attributes, have been largely studied as an imaging modality for these applications. However, the number of points required to faithfully represent surfaces is very high, requiring effective compression. Lossy compression techniques achieve high compression ratios by discarding part of the information from the source data, which can affect the visual quality of the decoded content. This impact can be accurately measured in subjective quality assessment experiments, which are, however, expensive and time-consuming. Objective metrics are alternative approaches to the estimation of visual quality, being designed to have a high correlation with subjective scores.
Despite the efficiency of modern media compression techniques, the demand for data storage has been rapidly growing. Current data centers consume energy to maintain the data and have a limited lifespan. In recent years, DNA-based storage has been studied as an alternative technique, with different methods being proposed to encode data into the nucleotides of DNA molecules. This technology showcases interesting features, such as very low energy consumption, high volumetric information density, and the capacity to last for centuries. However, the DNA storage channel also presents various challenges, in particular the presence of noise due to errors added during the synthesis and sequencing process.
This thesis addresses the issue of point cloud compression by proposing several coding algorithms based on neural networks. The studies proposed for static geometry compression evaluate how techniques such as predictive, residual, and autoregressive entropy coding can be adapted to learning-based models. Moreover, a comparative study of the effect of two types of convolutional networks on compression performance is conducted. An architecture for color compression combining graph Fourier transforms with a learned entropy model is also proposed and compared to conventional standards. Finally, a conditional coding algorithm leveraging temporal redundancy in dynamic sequences is presented.
This thesis also describes three subjective quality assessment experiments that evaluated point clouds compressed with both conventional and learning-based solutions. These experiments also consider other aspects related to subjective assessment, such as best practices for crowdsourcing-based evaluations, the use of immersive displays, and the bitrate allocation between color and geometry employed during compression. The obtained opinion scores were used to benchmark the performance of several state-of-the-art objective metrics, producing valuable insights regarding the features that better correlate with human perception. Moreover,