Image Compression, Vision, and Computer Graphics

fields, are converging rapidly, thus enabling several interesting new
networked media applications. In this talk, I will review our work at the
intersection of image compression and graphics.

The first part of the talk will discuss the compression of 4-d
light-fields. We compare the encoding of view-dependent texture maps using
4-d wavelets against a hierarchical image-domain light-field coder and
show, why image-domain encoding is (usually) superior to texture-map encoding.

The second part of the talk presents results of model-based compression for
talking-head video sequences. I will show how we model and automatically
track facial expressions, how these techniques enable photorealistic
avatars, and how we can incorporate synthetic video in standardized
motion-compensated hybrid coders to significantly improve their compression
efficiency.

Our work shows that 3-d geometry models can indeed help to compress images,
if many views of the same 3-d object or scene shall be compressed. This has
long been suspected. In each case, accurate vision algorithms for 3-d
reconstruction are a key component of the system.