Image Segmentation via Spectral Clustering with Density Sensitive Similarity Function

File structure:

├── data/
|   └── cifar-10-python.tar.gz
├── demo.py
├── DSSC.py
├── evaluation.py
├── kmeans.py
├── ncut.py
├── preprocessing.py
├── test0.png
├── test1.png
├── test2.png
├── utils.py

How to run it:

Reproduce the result in report by:

python demo.py 

The default label is plane.

You can try different labels by

python demo.py --label='your label'

The supported labels in demo are plane, horse and deer.

Reproduce Segmentation result by three different models

kmeans:

python kmeans.py

The default label is using kmeans 3d ++ model.

You can try kmeans 5d ++ model by

python kmeans.py --dim=5

ncut:

python ncut.py

DSSC:

python DSSC.py

Evaluation:

python evaluation.py

The model is evaluated by three different metrics -- Accuracy, F-score, NMI. The evaluation evaluates three models' performance on horse, deer and plane. We manually label these pictures as test*.png for evaluation.

Note that: The table below is ran under 10 test pictures, in submission, we only use 3 test images as example.

Algorithm	ACC	F-score	NMI
Kmeans++ 3D	0.69	0.72	0.54
Kmeans++ 5d	0.63	0.57	0.48
Ncut	0.71	0.66	0.52
DSSC	0.76	0.75	0.62