Joint re-identification and damage detection in the insurance domain: training from synthetic data

This thesis is part of a project that is concerned with the study of object re-identification and damage detection applied to images of personal belongings, in particular bikes. Object re-identification (ReID) is a computer vision based technology which aims at associating a particular object across different scenes and camera views, in order to assess whether a specific object is present in an image or video sequence regardless of the background or the object angle and position. Damage detection (DT) is the automatic process of identifying the presence of a damage on an object, and eventually assess the location and the type of the damage. The first focus point of this thesis is the design of a computer graphics pipeline for the semi-automatic creation of synthetic data to be used for training of a Transformer-based neural network, to address the lack of training data regarding damaged bikes. In particular, through the use of computer graphics we are able to generate images of the same bike before and after a particular damage. The final dataset is a collection of images of different bikes, with randomly generated appearances, damages of various types and missing parts, placed in multiple real-world environments. The second focus point of the thesis is Domain Adaptation: when training a neural network with mostly synthetic data, its behavior when faced with the real world data it is supposed to work with can be unpredictable. Domain adaptation is a family of techniques that aim at reducing the domain gap between synthetic data and real data, so that the network cannot perceive any difference between the two. During the research different configurations, both internal and external to the network, along with the implementation of multiple analysis tools, have been employed to achieve a deep understanding of the impact that a synthetic-based training has on the network and to improve its performance when working with real world images.