# YOLO-FaceV2: A Scale and Occlusion Aware Face Detector

**Authors:** Ziping Yu, Hongbo Huang, Weijun Chen, Yongxin Su, Yahui Liu, Xiuying Wang  
**arXiv:** [2208.02019](https://arxiv.org/abs/2208.02019)  
**Date:** 2022-08-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.02019v1/

## Key points

- Significant progress in face detection algorithms based on deep learning
- Two types of algorithms: two-stage detectors (Faster R-CNN) and one-stage detectors (YOLO)
- One-stage detectors gaining popularity due to better balance between accuracy and speed
- Introduction of YOLO-FaceV2, a real-time face detector based on YOLOv5
- Receptive Field Enhancement (RFE) module to enhance detection of small faces
- NWD Loss to compensate for inaccuracies caused by slight deviations in object localization
- SEAM attention module to handle face occlusion and improve detection accuracy
- Repulsion Loss to further improve detection performance in occlusion scenarios
- Slide weight function to address sample imbalance between easy and hard samples during training
- Anchors designed for improved localization accuracy using effective receptive field information
- Evaluation on WiderFace dataset shows YOLO-FaceV2 outperforms YOLO and its variants across all subsets (easy, medium, and hard)
- YOLO-FaceV2 achieves state-of-the-art performance in real-time face detection.

## Summary

In recent years, there has been significant progress in face detection algorithms based on deep learning. These algorithms can be categorized into two types: two-stage detectors like Faster R-CNN and one-stage detectors like YOLO. One-stage detectors have gained popularity due to their better balance between accuracy and speed, making them suitable for various applications. This paper introduces YOLO-FaceV2, a real-time face detector based on the one-stage detector YOLOv5. The authors propose several enhancements to improve the performance of the detector. Firstly, they design a Receptive Field Enhancement (RFE) module called RFE to enhance the receptive field of small faces. This module addresses the challenge of accurately detecting small faces by improving the model's ability to capture fine-grained details. To address the sensitivity of Intersection over Union (IoU) to location deviation in tiny objects, the authors introduce NWD Loss which helps compensate for inaccuracies caused by slight deviations in object localization. The paper also tackles the issue of face occlusion by introducing an attention module named SEAM which focuses on regions with occluded faces allowing for more accurate detection even when faces are partially obscured. Additionally, Repulsion Loss is introduced to further improve detection performance in occlusion scenarios. To address sample imbalance between easy and hard samples a weight function called Slide is utilized which assigns higher weights to harder samples during training ensuring that the model learns from challenging examples effectively. Furthermore, leveraging information from effective receptive field anchors are designed for improved localization accuracy. The proposed YOLO-FaceV2 is evaluated on the WiderFace dataset and outperforms YOLO and its variants across all subsets (easy, medium, and hard). The experimental results demonstrate that YOLO-FaceV2 achieves state-of-the-art performance in real-time face detection.