Granular Computing and Sequential Analysis of Deep Embeddings in Fast Still-to-Video Face Recognition

This paper is focused on still-to-video face recognition with large number of subjects based on computation of distances between high-dimensional embeddings extracted using deep convolution neural networks. We propose to utilize granular structures and sequentially process granular representations of all frames of the input video. The coarse-grained granules include only low number of the first principal components of deep embeddings. The representation of each frame at finer granularity levels is matched with the representations of photos of only those individuals, for whom the decision at previous levels was reliable. The reliability is checked by thresholding the ratio of distance between reference instance and input frame to the minimal distance. As a result, the photos of all unreliable individuals are not examined anymore for a particular frame at the next levels with finer granularity. Decisions for all frames are united into a candidate set of identities, and the maximal a-posterior final decision is chosen. The experimental study with the LFW, YTF and IJB-A datasets and the state-of-the-art deep embeddings demonstrated that the proposed approach is 2–10 times faster than conventional methods.