Towards Flops-constrained Face Recognition

TL;DR

This paper presents a flops-constrained face recognition approach that combines a searched network architecture, novel loss and frame aggregation methods, achieving state-of-the-art results within a 30 GFlops limit.

Contribution

The paper introduces Efficient PolyFace, ArcNegFace, and QAN++, novel methods optimized for low-flops face recognition tasks.

Findings

Achieved 94.198% accuracy at 1e-8 on deepglint-large.

Achieved 72.981% accuracy at 1e-4 on iQiyi-large.

Solutions operate within a 30 GFlops budget.

Abstract

Large scale face recognition is challenging especially when the computational budget is limited. Given a \textit{flops} upper bound, the key is to find the optimal neural network architecture and optimization method. In this article, we briefly introduce the solutions of team 'trojans' for the ICCV19 - Lightweight Face Recognition Challenge~\cite{lfr}. The challenge requires each submission to be one single model with the computational budget no higher than 30 GFlops. We introduce a searched network architecture `Efficient PolyFace' based on the Flops constraint, a novel loss function `ArcNegFace', a novel frame aggregation method `QAN++', together with a bag of useful tricks in our implementation (augmentations, regular face, label smoothing, anchor finetuning, etc.). Our basic model, `Efficient PolyFace', takes 28.25 Gflops for the `deepglint-large' image-based track, and the `PolyFace+QAN++' solution takes 24.12 Gflops for the `iQiyi-large' video-based track. These two solutions achieve 94.198\% @ 1e-8 and 72.981\% @ 1e-4 in the two tracks respectively, which are the state-of-the-art results.