PORTALE DELLA DIDATTICA

Ricerca CERCA
  KEYWORD

Using generative AI to create annotated datasets for wet damage identification

keywords GENERATIVE AI, ANNOTATED DATASETS, INSTANCE SEGMEN

Reference persons ANDREA BOTTINO

Research Groups DAUIN - GR-02 - COMPUTER GRAPHIC AND VISION GROUP - CGVG

Thesis type RESEARCH / EXPERIMENTAL

Description This work focuses on the automatic detection of wet damages from images. A major challenge in this work is the lack of annotated datasets large enough to effectively train instance segmentation algorithms. The core idea of this thesis proposal is to use the capabilities of generative AI to create a synthetic dataset of annotated images. Using a small set of existing annotated images, the work aims to develop a generative AI approach (GenAI) that can extrapolate and generate a comprehensive dataset that simulates different scenarios and conditions of wet damage. This dataset will then be used to train robust instance segmentation algorithms to improve their accuracy and effectiveness in real-world applications.

The expected outcome of this work is twofold. First, to successfully demonstrate the feasibility of using generative AI to create large, diverse and reliable annotated datasets from a minimal number of real annotated images. Second, to evaluate the performance of instance segmentation algorithms trained on these synthetic datasets.

Possibility of expense reimbursement (with a research grant).

See also  https://areeweb.polito.it/ricerca/cgvg/thesis.html

Required skills Machine Learning, Deep Learning, Python


Deadline 15/01/2025      PROPONI LA TUA CANDIDATURA