Visual Similarity Detection Based on Latent Representations

Kurzfassung

In this thesis we develop and evaluate multiple scalable solutions for the task of image similarity detection as part of an automated testing system for the rendering pipeline of the game Space Engineers by GoodAI. We implement image similarity detectors based on comparison of compact representations of the input images generated by three self-supervised representation learning architectures: Multi-Path Augmented AutoEncoder by Sundermeyer et al. [12] and two kinds of Siamese networks partially based on the first architecture. We demonstrate the applicability of the aforementioned architectures in the given setting of synthetic training data and existing domain gap between the training and the application domains and evaluate their ability to produce latent representations which meet the requirements of robust generalization and invariance to variations in background, lighting, texturing and object’s pose within the field of view. We demonstrate the benefits of the multi-path architecture for the descriptiveness of latent representations with respect to the appearance features of the objects in the images as opposed to the geometric features examined in the original paper [12].