Research Interests

Computational Vision

My current research interests span the three levels of machine vision, namely low, intermediate and high level vision. A brief description of my interests and opinions in each of these areas is given below.

Low-Level Vision

This level is concerned with the measurement of image primitives. I have done work in the use of phase information for visual motion estimation, stereo disparity, image contours, and view-based object recognition. (See the papers: Fleet and Jepson, 1993, Fleet, Jepson and Jenkin, 1991 Jepson and Fleet, 1991.) I am still involved with some colour-constancy work.

Opinions: While we are still far from an `optimal' set of low-level operations, the techniques available now for the measurement of image primitives (from across the field of machine vision, not just the ones I have worked on) are sufficient to support the subsequent levels of processing. The critical remaining issue appears to me to be how to squeeze the various measurement operations into real-time hardware, which is beyond my scope of expertise.

Intermediate-Level Vision

This level is in general concerned with the the identification and representation of `structure' in the low-level measurements. I am working on general purpose techniques for this based on layered mixtures of robust local models. Applications include:

Optical Flow:
- Ju, Black and Jepson, 1996, Skin and Bones: Multi-layer, locally affine, optical flow and regularization with transparency.
- Black and Jepson, 1996a, Estimating multiple independent motions in segmented images using parametric models with local deformations.
- Jepson and Black, 1993, Mixture models for optical flow.
Stereo Disparity:
- Jepson and Jenkin, 1994, Detecting floor anomalies.
Image Contour
Grey-level, Colour, and Depth Data:
- Jepson and Black, 1996, Mixture models for image representation.
View-Based Tracking:
- Black and Jepson, 1996b, EigenTracking: Robust matching and tracking of articulated objects using a view-based representation.

Related work includes the estimation of 3D motion (see Jepson and Heeger, 1991), and the simultaneous identification of objects moving independently in 3D (see MacLean, Jepson and Frecker, 1994).

Opinions: I believe we, as a field, can make a qualitative improvement in the robustness and accuracy of intermediate level approaches through the use of mixtures of robust local models and related techniques. Moreover, I believe this will take us to reasonably general purpose intermediate-level systems. There are many interesting remaining problems here, including:

on-line learning of specific forms for layered models,
data fusion across different modalities,
the use of (possibly structured) prior information,
the use of high-level control information,
stochastic or sparse sampling techniques for algorithm speed-ups.

Real-time implementations of these techniques are feasible, but in the near-term will be limited to simplified cases.

High-Level Vision

This level is in general concerned with scene understanding. As one part of this, I am working on particular situations in which perceptual inferences can be expected to be reliable. This work rests on the presence of particular `modal' structure in the domain (see Jepson, Richards and Knill, 1996, Richards, Jepson and Feldman, 1996). We are currently pursuing the applications of this sort of reasoning to the understanding of simple motion events (see Mann, Jepson and Siskind, 1996).

Opinions: Getting this form of basic high level reasoning component coupled up with the previous levels, at least in simple domains, is on the horizon.

Other information :

Papers available on-line
Software available
Address
Home page