Adaptive visual sampling

Abstract

Various visual tasks may be analysed in the context of sampling from the visual field. In visual psychophysics, human visual sampling strategies have often been shown at a high-level to be driven by various information and resource related factors such as the limited capacity of the human cognitive system, the quality of information gathered, its relevance in context and the associated efficiency of recovering it. At a lower-level, we interpret many computer vision tasks to be rooted in similar notions of contextually-relevant, dynamic sampling strategies which are geared towards the filtering of pixel samples to perform reliable object association. In the context of object tracking, the reliability of such endeavours is fundamentally rooted in the continuing relevance of object models used for such filtering, a requirement complicated by realworld conditions such as dynamic lighting that inconveniently and frequently cause their rapid obsolescence. In the context of recognition, performance can be hindered by the lack of learned context-dependent strategies that satisfactorily filter out samples that are irrelevant or blunt the potency of models used for discrimination. In this thesis we interpret the problems of visual tracking and recognition in terms of dynamic spatial and featural sampling strategies and, in this vein, present three frameworks that build on previous methods to provide a more flexible and effective approach. Firstly, we propose an adaptive spatial sampling strategy framework to maintain statistical object models for real-time robust tracking under changing lighting conditions. We employ colour features in experiments to demonstrate its effectiveness. The framework consists of five parts: (a) Gaussian mixture models for semi-parametric modelling of the colour distributions of multicolour objects; (b) a constructive algorithm that uses cross-validation for automatically determining the number of components for a Gaussian mixture given a sample set of object colours; (c) a sampling strategy for performing fast tracking using colour models; (d) a Bayesian formulation enabling models of object and the environment to be employed together in filtering samples by discrimination; and (e) a selectively-adaptive mechanism to enable colour models to cope with changing conditions and permit more robust tracking. Secondly, we extend the concept to an adaptive spatial and featural sampling strategy to deal with very difficult conditions such as small target objects in cluttered environments undergoing severe lighting fluctuations and extreme occlusions. This builds on previous work on dynamic feature selection during tracking by reducing redundancy in features selected at each stage as well as more naturally balancing short-term and long-term evidence, the latter to facilitate model rigidity under sharp, temporary changes such as occlusion whilst permitting model flexibility under slower, long-term changes such as varying lighting conditions. This framework consists of two parts: (a) Attribute-based Feature Ranking (AFR) which combines two attribute measures; discriminability and independence to other features; and (b) Multiple Selectively-adaptive Feature Models (MSFM) which involves maintaining a dynamic feature reference of target object appearance. We call this framework Adaptive Multi-feature Association (AMA). Finally, we present an adaptive spatial and featural sampling strategy that extends established Local Binary Pattern (LBP) methods and overcomes many severe limitations of the traditional approach such as limited spatial support, restricted sample sets and ad hoc joint and disjoint statistical distributions that may fail to capture important structure. Our framework enables more compact, descriptive LBP type models to be constructed which may be employed in conjunction with many existing LBP techniques to improve their performance without modification. The framework consists of two parts: (a) a new LBP-type model known as Multiscale Selected Local Binary Features (MSLBF); and (b) a novel binary feature selection algorithm called Binary Histogram Intersection Minimisation (BHIM) which is shown to be more powerful than established methods used for binary feature selection such as Conditional Mutual Information Maximisation (CMIM) and AdaBoost.