2D clustering with computer vision
Clustering algorithms often struggle with edge cases and do typically not have an induced bias to continue shapes. The idea of this thesis is to get better clustering results on 2D data (either directly, or indirectly through TSNE/UMAP) by exploiting modern instance segmentation models and finetuning them on the clustering task.
The thesis would include the creation of datasets where many traditional clustering methods fail, as well as finetuning computer vision models for the clustering task.
Prerequisites:
- Solid programming skills in python
- At least basic knowledge in machine learning and deep learning
