Abstract : A decade ago, 3D content was restricted to a few applications – mainly games, 3D graphics andscientific simulations. Nowadays, thanks to the development cheap and efficient specialized renderingdevices, 3D objects are ubiquitous. Virtually all devices with a display – from a large visualizationclusters to smart phones – now integrate 3D rendering capabilities. Therefore, 3D applications arenow far more diverse than a few years ago, and include for example real-time virtual and augmentedreality, as well as 3D virtual worlds. In this context, there is an ever increasing need for efficient toolsto transmit and visualize 3D content.In addition, the size of 3D meshes always increases with accuracy of representation. On one hand,recent 3D scanners are able to digitalize real-world objects with a precision of a few micrometers, andgenerate meshes with several hundred million elements. On the other hand, numerical simulationsalways require finer meshes for better accuracy, and massively parallel simulation methods now generatemeshes with billions of elements. In this context, 3D data compression – in particular 3D meshcompression – services are of strategic importance.The previous decade has seen the development of many efficient methods for encoding polygonalmeshes. However, these techniques are no longer adapted to the current context, because they supposethat encoding and decoding are symmetric processes that take place on the same kind of hardware.In contrast, remote 3D content will typically be created, compressed and served by high-performancemachines, while exploitation (e.g. visualization) will be carried out remotely on smaller – possiblyhand held – devices that cannot handle large meshes as a whole. This makes mesh compression anintrinsically asymmetric process.Our objective in this dissertation is to address the compression of these large meshes. In particularwe study random-accessible compression schemes, that consider mesh compression as an asymmetricproblem where the compressor is an off-line process and has access to a large amount of resources,while decompression is a time-critical process with limited resources. We design such a compressionscheme and apply it to interactive visualization.In addition, we propose a streaming compression algorithm that targets the very large hexahedralmeshes that are common in the context of scientific numerical simulation. Using this scheme, we areable to compress meshes of 50 million hexahedra in less than two minutes using a few megabytes ofmemory.Independently from these two specific algorithms, we develop a generic theoretical framework toaddress mesh geometry compression. This framework can be used to derive geometry compressionschemes for any mesh compression algorithm based on a predictive paradigm – which is the case of thelarge majority of compression schemes. Using this framework, we derive new geometry compressionschemes that are compatible with existing mesh compression algorithms but improve compressionratios – by approximately 9% on average. We also prove the optimality of some other schemes underusual smoothness assumptions.