Improving precision and recall of face recognition in SIPP with combination of modified mean search and LSH

TL;DR

This paper introduces a novel combination of modified mean search and LSH for SIPP face recognition, significantly improving precision and recall without retraining DNN models, demonstrated on MSceleb and CASIA datasets.

Contribution

It proposes a unique rule-based combination of modified mean search and LSH, along with a SVD-based augmentation, to enhance face recognition performance without retraining DNNs.

Findings

Coverage increased from 13.39% to over 47% at high precision levels.

Achieved top-10 ranking in MSceleb challenge without fine-tuning.

Consistent improvements on CASIA WebFace dataset.

Abstract

Although face recognition has been improved much as the development of Deep Neural Networks, SIPP(Single Image Per Person) problem in face recognition has not been better solved, especially in practical applications where searching over complicated database. In this paper, a combination of modified mean search and LSH method would be introduced orderly to improve the precision and recall of SIPP face recognition without retrain of the DNN model. First, a modified SVD based augmentation method would be introduced to get more intra-class variations even for person with only one image. Second, an unique rule based combination of modified mean search and LSH method was proposed the first time to help get the most similar personID in a complicated dataset, and some theoretical explaining followed. Third, we would like to emphasize, no need to retrain of the DNN model and would easy to be extended without much efforts. We do some practical testing in competition of Msceleb challenge-2 2017 which was hold by Microsoft Research, great improvement of coverage from 13.39% to 19.25%, 29.94%, 42.11%, 47.52% at precision 99%(P99) would be shown latter, coverage reach 94.2% and 100% at precision 97%(P97) and 95%(P95) respectively. As far as we known, this is the only paper who do not fine-tuning on competition dataset and ranked top-10. A similar test on CASIA WebFace dataset also demonstrated the same improvements on both precision and recall.