Novel feature construction technique for detecting anomalous faces and evaluating style transfer methods

Abstract

In this thesis, we focus on how careful design and evaluation of deep-learned features are still necessary like hand-crafted features for computer vision tasks. We demonstrate this in two different domain problems – Anomaly Detection and Style Transfer. We present feature aggregation techniques and also quantitative evaluation procedure for these tasks. For anomaly detection, we propose a novel facial anomaly detection task, where we demonstrate a feature extraction procedure using a specially trained autoencoder for detecting anomalous faces without seeing any example anomalies during training. We built a new dataset of anomalous faces and typical faces for evaluating the proposed framework that beats many standard baselines. For style transfer, we developed the first quantitative evaluation procedure for evaluating existing style transfer methods using an effectiveness and coherence metric to measure how effectively a style has transferred without distorting object boundaries much. Doing so, helped us to design better features for extracting style using cross-layer gram matrices instead of popularly adopted within later gram matrices. Both works signify understanding features and their careful design are still crucial in building state of the art computer vision algorithms.