Clustering approach where the consistency is preserved.
Using approach in the paper The The Chosen One: Consistent Characters in Text-to-Image Diffusion Models , I augmented clustering approach to identify consistent image clusters for a given dataset.
consistency refers to finding a shared representation across all the other given instances.
Mathematically, we can represent cohesion as below,
For the above image, the most cohesive cluster would be the cluster depicting black dots as its instances are more closer to its cluster centroid(higher consistency).
If the clustering approach works well, we can find the optimal set clusters that each of them has unique set of patterns. Then we can find the most cohesive cluster, aka the cluster that shows the most consistency across all the instances(features are same and consistent).
Install necessary packages,
pip install -r requirements.txt# Edit the config files
- edit main_config file
- set the hyperparametes in preferred algorithm
# Run the script
python clustering.py- Implemented kmeans++
- Implemented DBSCAN
- Implemented BIRCH
- Testing with Graph Clustering
- Top-k retrieval
- Support for rotational invariance of the images(see below)
- Overall results are fine, there are some edge cases that need to handle seperately.
- Some clusters might have mixed-images(probably due to either extracted features or clustering algorithm in used.)
- Sensitive to the rotation of the image, needs to be addressed (Make this as Rotational Invariant Method).
- Enhance the feature extraction process: Incoporate more features that make them unique!
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
Please make sure to update tests as appropriate.