Efficient learning methods for high dimensional visual data
Marcus Caixing Chen · 2015
High dimensional visual data, derived from images or videos, is ubiquitous as advance camera technologies enable more measurements per sample to be captured.Increasingly sophisticated visual data representations further contribute to an increase in data dimensionality.To facilitate high level visual analytic tasks, this thesis focuses on three areas of high dimensional data processing, namely direct graph embedding for sample class or cluster prediction, video tracking for temporal information extraction, and spatial segmentation for a compact representation of high resolution images.To address the challenges of irrelevant, noisy, and highly correlational dimensions, a novel unified framework is proposed to simultaneously perform graph embedding and feature selection.This framework enables an efficient extraction of linear data intrinsic structures, which are low dimensional and robust to both noisiness in dimensions and outlier samples.This framework is computationally efficient and flexible to incorporate various data prior properties such as smoothness, sparsity, and locality.In video analysis, efficient learning of high dimensional visual data often requires modeling of temporal evolution of object appearance and motion.Instead of analyzing all the visual data, object level temporal information can be extracted via visual tracking for more efficient learning.For a long video sequence, the object appearance will change due to variations in its poses and orientation, illumination, and occlusion.To track both the object appearance and position, it is necessary to have a robust tracker with an adaptive object appearance update.We propose a generative model to address the dual uncertainties in both the target positions and appearance simultaneously.A diffusion process on a Riemannian manifold allows a geodesic evolution of the target appearance.Spatially, object segmentation can significantly simplify visual learning by grouping many pixels into a meaningful representation.However useful, object segmentation remains unsolved.We propose a novel object co-segmentation framework to learn object segmentation using a large image set.By leveraging on good segmentation results on the simplest images, we can propagate this to more and more complex images.i Over the past five years, I have received support and encouragement from a great number of individuals.Professor Cham Tat Jen has been a very encouraging mentor.At the beginning of my Ph.D. studies, Professor Cham made numerous trips to my company for discussions.His guidance and encouragement have made this a thoughtful and rewarding journey.