Thesis Description & Objectives
Object detection methods strongly rely on supervised deep learning approaches that
require large amounts of labeled data. Data augmentation techniques can be employed
to increase dataset size. Among these techniques, the generation of fully synthetic data
is particularly promising, as it provides automatically labeled data, thereby saving time
and effort.
However, these approaches present several drawbacks, as they often require non-trivial
processes, such as 2D image-to-3D model conversion or photorealistic rendering.
Furthermore, these challenges are amplified in industrial environments, which often
involve highly specific assets and objects that are not easily found in publicly available
datasets.
Therefore, this thesis aims to address these limitations by implementing an effective
pipeline that enables human operators to easily generate synthetic datasets for object
detection tasks in industrial scenarios. The thesis will first explore recent approaches
(e.g., 2D-to-3D conversion, Generative Adversarial Networks, diffusion models) and will
subsequently propose a complete pipeline, which will be evaluated on a real industrial
use case.
Examples of similar/useful ideas:
• https://www.youtube.com/watch?v=HHzNIh72B_Y
• https://www.youtube.com/watch?v=iuDv3TS-xQA
• https://www.youtube.com/watch?v=dxlyCPGCvy8
• https://www.youtube.com/watch?v=pJfHiVo8GUg
• https://www.youtube.com/watch?v=BuL7RroTG7s
