Optimization for Greedy Non-maximum Suppression Based on Multi-task Convolutional Neural Network

Qingqing Hong,Qian Fan,Lifeng Zhang

doi:10.12792/jiiae.8.39

Abstract

Non-maximum suppression (NMS) is an essential part of the face detection pipeline based on a convolutional neural network (CNN). The typical approach for NMS used by face detection is a greedy, locally optimal strategy, which is to localize objects from a set of candidate locations. However, NMS still has some shortcomings, such as sometimes the detection box has no relationship with a high classification score, which leads to misjudging face localization during NMS. In this paper, we observed that the NMS implemented on the multi-task convolutional neural network (MTCNN) is a cascaded network and the enhanced NMS based on MTCNN to achieve high performance during face detection and alignment. We employ WIDER FACE as the test dataset to evaluate our proposal. The precision and recall curves are drawn when using three subsets at different thresholds. The result shows that the proposed approach can perform better performance than traditional NMS.

Highlights

Face detection and alignment has been widely applied in computer vision to generate bounding boxes for identifying frontal face and assign them classification scores
We introduce a new framework represented by multi-task cascaded convolutional networks (MTCNN) to integrate face detection and alignment proposed by Kaipeng Zhang et al.[2]
This paper adopted a new Non-maximum suppression (NMS) algorithm to improve the performance of MTCNN

Summary

Introduction

Face detection and alignment has been widely applied in computer vision to generate bounding boxes for identifying frontal face and assign them classification scores. Even if Chen et al.[1] presented a new state-of-art approach for face detection and alignment. We introduce a new framework represented by multi-task cascaded convolutional networks (MTCNN) to integrate face detection and alignment proposed by Kaipeng Zhang et al.[2]. The advantages of MTCNN is described as follows: First, MTCNN adopted a new cascade and lightweight CNN framework to combine face detection and alignment for achieving excellent realtime performance. The characteristics of greedy NMS is to make a hard decision and set a fixed threshold to decide the range of suppression. In this case, face misjudging will be occurred and the average precision of the entire network may be dropped.

Related Work

Proposal

Experiment

Findings

Conclusions