Proxy Embeddings For Face Identification Among Multi-Pose Templates

Many of a large scale face identification systems operates on databases containing images showing heads in multiple poses (from frontal to full profiles). However, as it was shown in the paper, off-the-shelf methods are not able to take advantage of this particular data structure. The main idea behind our work was to adapt the methods proposed for multi-view and semi-3D objects classification to the multi-pose face recognition problem. The proposed approach involves neural network training with proxy embeddings and building the gallery templates out of aggregated samples. A benchmark testing scenario is proposed for the purpose of the problem, which is based on the linked gallery and probes databases. The gallery database consists of multipose face images taken under controlled conditions, and the probes database contains samples of in-the-wild type. Both databases must be linked, having at least partially common labels. Two variants of the proposed training procedures were tested, namely, the neighbourhood component analysis with proxies (NCA-proxies) and the triplet margin loss with proxies (triplet-proxies). It is shown that the proposed methods perform better than models trained with cross-entropy loss and than off-the-shelf methods. Rank-1 accuracy was improved from 48.82% for off-the-shelf baseline to 86.86% for NCA-proxies. In addition, transfer of proxy points between two independently trained models was discussed, similarly to hyper-parameters transfer methodology. Proxy embeddings transfer opens a possibility of training two domain-specific networks with respect to two datasets identification schema.