Visual object detection across different domains by solving self supervised tasks

Deep Neural Network models based on convolutions need a large dataset to be trained successfully. In the Computer Vision context this implies that a lot of labeled images have to be collected in order to obtain a model with good performances. A trained model is then often unusable when exploited in a visual domain which is different from the training one; moreover the data collection and labeling process can be physically or economically impossible in some visual domains. From these considerations the need to develop algorithms robust to visual domain shifts. ?? Self supervised tasks have shown a great potential as a strategy to learn useful features from unlabeled images. They can therefore be used in cross domain analysis as a method to obtain feature alignment between different domains. The purpose of this thesis is to study how self supervised tasks can be used to develop well performing visual object detection models in various cross domain analysis settings: Domain Generalization, Domain Adaptation and a new more general and challenging setting called One-Sample adaptation.