Master Assignment
[M] Aerial image segmentation with data-efficient adaptation of foundation models
Type: Master EE/CS/ITC
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Background:
Drones are increasingly being used in low- and middle-income countries to gain insights for disaster risk management, slum upgrading, nature-based solutions, and many other development issues (e.g. [1]). AI has the opportunity to automatically map features of interest in this imagery, yet the objects to be detected are often use-case specific and there is no labelled data of these objects to train supervised algorithms.
Objectives:
The objective of this topic is to consider how foundational segmentation models, such as the Segment Anything Model [2], can be used to speed up the discovery of interesting parts of the image (i.e. detect objects that are relevant for disaster risk management and sustainable development). This will contribute to a faster workflow for practitioners to leverage the advantages of AI for new applications.
The project aims at investigating techniques for generalization and domain adaptation using data-efficient approaches, f.i. learning with few shots [3] or augmenting computer vision algorithm with prior knowledge [4], in the context of aerial image analysis. It is a collaboration between the EEMCS and ITC Faculties.
[1] Monitoring household upgrading in unplanned settlements with unmanned aerial vehicles - https://doi.org/10.1016/j.jag.2020.102117
[2] Segment Anything - https://arxiv.org/abs/2304.02643
[3] Few-shot learning - https://en.wikipedia.org/wiki/One-shot_learning_(computer_vision)
[4] Data-efficient Large Scale Place Recognition with Graded Similarity Supervision - https://arxiv.org/abs/2303.11739