[Paper Review] Sample and Computation Redistribution for Efficient Face Detection

Jia Guo, Jiankang Deng+2 more|arXiv (Cornell University)|May 10, 2021

Face recognition and analysisComputer Science43 references68 citations

TL;DR

The paper introduces Sample Redistribution (SR) and Computation Redistribution (CR) to improve efficiency and accuracy in VGA-resolution face detection, achieving state-of-the-art trade-offs with SCRFD models.

ABSTRACT

Although tremendous strides have been made in uncontrolled face detection, efficient face detection with a low computation cost as well as high precision remains an open challenge. In this paper, we point out that training data sampling and computation distribution strategies are the keys to efficient and accurate face detection. Motivated by these observations, we introduce two simple but effective methods (1) Sample Redistribution (SR), which augments training samples for the most needed stages, based on the statistics of benchmark datasets; and (2) Computation Redistribution (CR), which reallocates the computation between the backbone, neck and head of the model, based on a meticulously defined search methodology. Extensive experiments conducted on WIDER FACE demonstrate the state-of-the-art efficiency-accuracy trade-off for the proposed \scrfd family across a wide range of compute regimes. In particular, \scrfdf{34} outperforms the best competitor, TinaFace, by $3.86\%$ (AP at hard set) while being more than \emph{3$ imes$ faster} on GPUs with VGA-resolution images. We also release our code to facilitate future research.

Motivation & Objective

• Motivate efficient face detection at VGA resolution with limited compute.
• Propose SR to increase training samples for shallow detector stages.
• Propose a two-step CR strategy to optimize computation across backbone, neck, and head.
• Demonstrate improved accuracy–efficiency trade-offs on WIDER FACE across compute regimes.

Proposed method

• Introduce Sample Redistribution (SR) by large-cropping to boost small-face training samples, especially for stride-8 features.
• Design a simplified computation search space for backbone, neck, and head based on fixed FLOP budgets.
• Perform two-step computation redistribution: (SCRFD 1) backbone-only optimization, (SCRFD 2) backbone+neck+head optimization under backbone guidance.
• Use empirical bootstrap to estimate promising computation distributions and reduce search space.
• Train models from scratch with VGA-only testing and compare to state-of-the-art detectors on WIDER FACE.

Experimental results

Research questions

• RQ1 How can training data sampling be redistributed to improve detection of small faces at VGA resolution?
• RQ2 How should computation be redistributed across backbone, neck, and head to maximize accuracy under a fixed FLOP budget?
• RQ3 What is the impact of a two-step CR strategy on detection performance and efficiency for face detectors?

Key findings

• SR significantly increases positive samples for stride-8, boosting hard-set AP (e.g., hard AP improves from 67.32 to 74.47 on baseline).
• Two-step CR (backbone-only then full detector) yields higher AP than single-step reallocation (e.g., CR@two-step achieves 71.37 vs 70.98).
• SCRFD-34GF achieves state-of-the-art efficiency–accuracy, surpassing TinaFace by 3.86 AP on hard, while being over 3× faster at VGA resolution.
• SCRFD models deliver substantial efficiency gains across regimes (0.5–34 GF) with favorable parameter counts (e.g., SCRFD-34GF: 9.80M).
• Under VGA testing, SCRFD-2.5GF reaches competitive precision with much lower compute (~2.5 GF) and inference time.
• The two redistribution strategies prove complementary and robust across multiple compute budgets.