LINK PREDICTION BASED FACE CLUSTERING USING VARIATIONAL ATTENTIONAL GRAPH AUTOENCODER
View/ Open
Date
2020-12-01Author
Verlekar, Harish Deepak
0000-0003-0420-1523
Metadata
Show full item recordAbstract
In this work, we address the problem of clustering faces according to their individual
identities present inherently in the dataset.The current clustering frameworks are either
based on some heuristic method or require labelled data for training the models,also
some of them make assumptions on data distribution or shape of the clusters.We have
framed the problem of forming clusters to that of link prediction on graphs and learn
how to do that in a completely unsupervised way by proposing to use Variational Graph
Autoencoders and use Graph Attentional Network as the Encoder. We call this network
as Variational Attentional Graph Autoencoder(VAGAE).Our framework is not based on
any assumptions of data distribution or shape of clusters and learns without any use of
labelled data.
We solve the problem in the following way, we first extract features from a feature
extractor which has been trained in an unsupervised way using Convolutional Adversarial
Autoencoders and have compared the results of it with a pre-trained Inception Resnet
feature extractor trained using FaceNet triplet loss algorithm. We then generate Instance
Subgraphs(ISG) for each instance by finding the K-Nearest neighbours for each instance
upto 2-hops and connect the edges to the instance only if it satisfies an approximate rank-
order distance threshold. We then pass the ISG’s to the Variational Graph Autoencoder
which uses a Graph Attention Network as an encoder to learn general graph structure
features from ISG’s and perform link prediction. We then transitively merge the link
prediction prediction results to form final clusters.
We evaluate our results on the first 50 faces from the Youtube Faces dataset and show
that the results are decent enough or even better in some metrics compared to the other
clustering methods.We have also evaluated our method on Link based Face Clustering
via Graph Convolution Network and have shown that we get decent performance on
data taken from same distribution for train-test, even though we train our model using
unsupervised learning.