Video Processing

We have been working for several years on biologically-inspired algorithms for encoding visual information.  This page presents some examples of algorithms running in real-time on early GPUs.  All of these demonstrations date back to around 2005/6, so are certainly dated, but they all use convolutional networks of our own design.

The video to the left illustrates an example of phase-invariant structure detection, in this case, somewhat confusingly, a face. Thus, it is ``phase-invariant face detector'' ! Phase is actually an important property in image structure, so invariance to it can be a double-edged sword. Indeed, recent work shows that combining phase-invariant with phase-selective appears to be best.

The video to the right illustrates a feature-based registration using RANSAC with a customised quadratic (in spatial variables) homography. This provides superior stabilisation in registering retinal image sequences taken pre- and pos- operatively. Cao, Ng & Bharath; see Annals of BMVA paper.

The video to the left illustrates a biologically-inspired approach to normalising over phase-invariant directional fields; this allows representations of in-plane spatial direction in a remarkably illumination tolerant way (note change in room lighting from live camera feed). Note that this is not edge-detection: it is a representation of orientation encoding.