floatydev/project577

Project of Course CS577

To run on your system

• Make a virtual enviroment using python 3.9 (using a virtualenv, not conda because we specific pip packages).
• Clone the repo

git clone https://github.com/FloatyDev/project577.git 

• Install dependencies using the following command.

pip install -r requirements.txt

Scripts

The main scripts of this project are mnist_finch.py and segment_finch.py

mnist_finch.py

Offers estimation performance of FINCH on MNIST_10K dataset (with plots). The metrics provided are:

• Normalized Mutual Information (NMI) Score
• Adjuster Rand Index (ARI)
• Silhouette Score
• Calinski-Harabasz Score
• Davies-Bouldin Score

After running the script results will be saved on mnist_results directory. The script evaluates FINCH with both image raw pixels as well as CNN classification features (From ResNet50). These weights are pre-extracted and saved on data directory of the project.

segment_finch.py

This script uses the foundation model Segment Aything to produce quality binary masks of the provided image (some test images are put on images directory). After that FINCH is applied to the features made by either Hu-Moment or HOG extracting methods to cluster the binary masks by shape. To run the script you have to specify the argument --algo that takes 2 values:

• hu (use Hu-Moments as the feature extraction method)
• hog (use hog as the feature extraction method, with predefined parameters) So for example if you want to run the scripot using hu-moments as the feature extraction method you execute as:

python segment_finch.py --algo hu

Masks from segment-anything model are saved on images directory. The cluster results are saved on sam_finch_results directory.

Note

Weights for SAM are downloaded automatically if they are not found.