Domain Adaption and Fusion for Domain-specific Traffic Sign Detection

Kurzfassung

Machine Learning models are becoming useful and high-accurate in a vast area of contexts, their use can be valuable especially in road safety applications. The quality and overall characteristics of the data used to train such models plays a key role to the efficacy and accuracy of Deep Learning algorithms. However, when the applicative domain is different from the training domain, the performance will inevitably degrade, causing the algorithm to be ineffective at runtime. This thesis aims to investigate the methods and processes of data modeling and domain adaptation that are necessary to avoid a performance loss in these instances, in particular for the case of traffic sign detection in a specific domain. This work shows that using a dataset composed of different data domains as training data can lead to an improvement of a deep learning model performance in a diverse and specific domain, and that the domain adaptation can also be used to successfully increase the accuracy of the fusion of two domains in a single dataset. In this particular case, the project aims to create a model capable of accurately recognizing traffic signs in the very specific domain of the North Rhine-Westphalia region (Germany), which is devoid of any ground truth. The work is first focused on adapting, through modeling, both the datasets and the hyperparameters of a deep learning model. Afterwards, a technique of domain adaptation through self-supervision is designed and experimented in both directions with respect to the training datasets, in order to obtain a good accuracy for the fusion of their domains through cross pseudo-labeling. Since in many cases no ground truth can be used to evaluate the model performance change in such a particular case, a set of pseudo-metrics has also been designed, both to measure the efficacy of the domain fusion and the object detection accuracy on the NRW target domain. It is shown that the our domain adaptation technique yields a better generalization on the target domain while keeping almost the same accuracy on the source domain; The application of semi-supervised learning through pseudo-labeling is also experimented, however the non-perfection of the predicted annotations inevitably harm the performances of the CNN model when used as training data; this shows that processes of semi-supervised learning are still ineffective in complex cases like this.