Open Set Segmentation on Aerial Images
Survey about Open Set techniques, with focus on semantic segmentation methods (Anomaly Segmentation) and aerial settings (Open Set/Open World recognition).
Requirements
- M.Sc. in Machine Learning, Data Science, Computer Science, Mathematics, Telecommunications, or similar
- Knowledge of Python
- Software development skills
- Basic concepts of image processing
- Basic concepts of data science, concerning data analysis, processing and machine learning
Description
Detecting out-of-distribution examples is important for safety-critical machine learning applications, including aerial imagery. Tasks like land cover classification usually deal with a fixed set of classes, mapping every pixel or image to one of these categories, regardless of the underlying distribution. This is far from optimal especially for operational purposes, where models are trained on localized datasets and applied in the wild on a larger scale. However, OOD segmentation in aerial images in a relatively unexplored task, where benchmarks are not well defined and solutions are still an open problem. Therefore, the purpose of this work is to provide a complete and exhaustive survey about Open Set techniques, with focus on semantic segmentation methods (Anomaly Segmentation) and aerial settings (Open Set/Open World recognition). The project will comprise several steps, from the analysis of the literature and current state-of-the-art approaches, to the implementation and evaluation of different methods on benchmark datasets. The thesis will include the following activities: (i) data identification, and acquisition if required (ii) data analysis and pre-processing (iii) predictive model development, leveraging on machine learning and deep learning techniques (PyTorch, Keras, etc..), (iv) performance evaluation, (v) result analysis and visualization.