The coding ecology: image coding via competition among experts

We consider the image-coding problem as a competitive ecology of specialists, each vying for the task of coding a portion of an image. Each specialist, or expert, derives its competitive advantage from its ability to concisely describe an underlying visual event (e.g., shadows, motion, occluding objects). In this paper, the metaphor of an auction informs the design of a predictive coder. Experts locate candidate regions of support, characterize the activity within, and submit their proposals in the form of bids passed along to an auctioneer. We describe the designs for six such experts, and examine candidate strategies for the decision unit (the auctioneer). We define protocols for comparing the bids, and compare overall coding performance using several examples.