Modeling and efficient optimization for object-based scalability and some related problems

MPEG-4 is the first visual coding standard that allows coding of scenes as a collection of individual audio-visual objects. We present mathematical formulations for modeling object-based scalability and some functionalities that it brings with it. Our goal is to study algorithms that aid in semi-automating the authoring and subsequent selective addition/dropping of objects from a scene to provide content scalability. We start with a simplistic model for object-based scalability using the "knapsack problem"--a problem for which the optimal object set can be found using known schemes such as dynamic programming, the branch and bound method and approximation algorithms. The above formulation is then generalized to model authoring or multiplexing of scalable objects (e.g., objects encoded at various target bit-rates) using the "multiple choice knapsack problem." We relate this model to several problems that arise in video coding, the most prominent of these being the bit allocation problem. Unlike previous approaches to solve the operational bit allocation problem using Lagrangean relaxation, we discuss an algorithm that solves linear programming (LP) relaxation of this problem. We show that for this problem the duality gap for Lagrange and LP relaxations is exactly the same. The LP relaxation is solved using strong duality with dual descent--a procedure that can be completed in "linear" time. We show that there can be at most two fractional variables in the optimal primal solution and therefore this relaxation can be justified for many practical applications. This work reduces problem complexity, guarantees similar performance, is slightly more generic, and provides an alternate LP-duality based proof for earlier work by Shoham and Gersho (1988). In addition, we show how additional constraints may be added to impose inter-dependencies among objects in a presentation and discuss how object aggregation can be exploited in reducing problem complexity. The marginal analysis approach of Fox (1966) is suggested as a method of re-allocation with incremental inputs. It helps in efficiently re-optimizing the allocation when a system has user interactivity, appearing or disappearing objects, time driven events, etc. Finally, we suggest that approximation algorithms for the multiple choice knapsack problem, which can be used to quantify complexity vs. quality tradeoff at the encoder in a tunable and universal way.