Clustering Images

Image Clustering

Embeddings which are learnt from convolutional Auto-encoder are used to cluster the images.

Since the dimensionality of Embeddings is big. We first reduce it by fast dimensionality reduction technique such as PCA.

This is required as T-SNE is much slower and would take lot of time and memory in clustering huge embeddings.

After that we use T-SNE (T-Stochastic Nearest Embedding) to reduce the dimensionality further.

Since these are unsupervised embeddings. Clustering might help us to find classes.

Clustering output.

The clusters are note quite clear as model used in very simple one.

T-SNE is takes time to converge and needs lot of tuning.

Also the embeddings can be learnt much better with pretrained models, etc.

The clustering script can be found here

It can be used with any arbitrary 2 dimensional embedding learnt using Auto-Encoders.