CelebrityFaces

Finds similar celebrities given the photo.

You can test the service here: http://95.213.170.235:50002

Example

Our algorithm:

Naive approach:

Algorithm

To build index:

1. Extract embeddings.
2. Cluster them with K-means algo into K=450 centroids.
3. Build inverse index: learn mapping from centroids to list of id of images.
4. Subtract centroid's center from corresponding images.
5. Build compressed descriptors: compute the sign of scalar products on 512 random directions and add small bias. This step reduces the size of descriptor from 128 to 8 64-bit numbers.

To process new image:

1. Load all the data in the memory from prev. algo.
2. Find top (K=10 out of 450) nearest centroids.
3. For each nearest centroid subtract centroid's center from the query.
4. Compute Hamming distance in the compressed descriptors space between the query and all the images in the nearest centroids for each of 8 64-bit numbers.
5. Take top N=10 images from each table and sort them using full descriptions by euclidean distance.
6. Output top 5 images.

This algo implemented in C++.