Large Scale Pattern Detection in Videos and Images from the Wild

Abstract

Pattern detection is a well-studied area of computer vision, but still current methods are unstable in images of poor quality. This thesis describes improvements over contemporary methods in the fast detection of unseen patterns in a large corpus of videos that vary tremendously in colour and texture definition, captured “in the wild” by mobile devices and surveillance cameras. We focus on three key areas of this broad subject; First, we identify consistency weaknesses in existing techniques of processing an image and it’s horizontally reflected (mirror) image. This is important in police investigations where subjects change their appearance to try to avoid recognition, and we propose that invariance to horizontal reflection should be more widely considered in image description and recognition tasks too. We observe online Deep Learning system behaviours in this respect, and provide a comprehensive assessment of 10 popular low level feature detectors. Second, we develop simple and fast algorithms that combine to provide memory- and processing-efficient feature matching. These involve static scene elimination in the presence of noise and on-screen time indicators, a blur-sensitive feature detection that finds a greater number of corresponding features in images of varying sharpness, and a combinatorial texture and colour feature matching algorithm that matches features when either attribute may be poorly defined. A comprehensive evaluation is given, showing some improvements over existing feature correspondence methods. Finally, we study random decision forests for pattern detection. A new method of indexing patterns in video sequences is devised and evaluated. We automatically label positive and negative image training data, reducing a task of unsupervised learning to one of supervised learning, and devise a node split function that is invariant to mirror reflection and rotation through 90 degree angles. A high dimensional vote accumulator encodes the hypothesis support, yielding implicit back-projection for pattern detection.