Batch of 2015, IIIT Delhi

An inferred social graph of the people in the Batch of 2015, IIIT Delhi.

When people get a photo clicked together, it can often be assumed that they share some social relation. This graph is derived from co-occurrence of individuals in the batch farewell photographs. Thus, if you got a photo clicked with another person in the farewell, you and that person share a link in this graph. This idea is leveraged to improve face recognition in group photos by almost 13% in our paper titled "Harnessing Social Context for Improved Face Recognition", accepted at International Conference on Biometrics 2015.

Each node in this undirected graph represents an individual present in the farewell album. The colour of the node denotes the "community" the person belongs to, detected using Louvain Modularity Scores. Each link indicates a co-occurrence of two individuals in one or more photos, with the weight storing the number of photos co-occurred in. In this force-layout representation, each node repels each other and links are weak geometric constraints. Think of links like rubber bands - higher weights imply less elasticity, shorter length and thicker links.

The graph rendered here is dynamic. Use scroll wheel to zoom, hover over a node to see its immediate neighbours and drag a node to play with the physics.

I got the photos from Tasveer (shameless plug), imported them into Picasa and manually labelled all the faces to get the ground truth. After writing the face tags to XMP metadata, I read the tags into python, created the graph using NetworkX and ran Louvain Community detection with a resolution of 0.5 to get 12 communities. Used Gephi to analyze the data and exported the graph to JSON, which is then fed to D3.js and visualized here.

If you are interested in doing more fancy stuff with the data, please feel free to do so. The data in JSON is available here.