无重叠子图限制的深度学习人脸识别聚类

需积分: 5 28 浏览量更新于2024-08-05 收藏 1.21MB PDF 举报

"《通过信心与连接性估计进行人脸聚类》(Learning to Cluster Faces via Confidence and Connectivity Estimation)是一篇关于深度学习在人脸识别领域的研究论文。随着人脸识别技术的广泛应用，尤其是在面部标注和检索等场景中，无监督或半监督的聚类方法变得尤为重要。然而，传统的聚类方法往往依赖于启发式步骤，并且需要大量的重叠子图，这在提高准确性和效率上存在局限。该论文提出了一个全新的、完全可学习的聚类框架，旨在解决这些问题。它不依赖于大量重叠子图，而是将聚类问题分解为两个子任务：一是通过图卷积网络GCN-V来估计每个节点（人脸）的信心度，二是通过另一图卷积网络GCN-E来评估边（人脸之间的相似度）的连接性。这种方法避免了手动设置阈值或复杂的预处理步骤，使得模型能够自主学习数据内在的结构和关系。 GCN-V利用节点的特征表示和邻域信息，学习每个节点属于特定簇的概率，而GCN-E则基于节点间的相互作用，动态调整边的权重，反映了不同人脸之间的相似程度。这两个网络的输出结合起来，形成了一个更为精确且高效的聚类策略。通过端到端的学习，模型能够自适应地捕捉数据中的模式，从而提升聚类性能。论文的核心创新在于将复杂的问题简化为可优化的网络架构，减少了人工干预，提高了聚类的稳定性和准确性。此外，这种方法还有望降低计算成本，使得大规模人脸数据的处理变得更加可行。这篇论文为无监督或弱监督下的人脸聚类提供了一个新的、有效的解决方案，有望推动该领域的发展。"

Learning to Cluster Faces via Conﬁdence and Connectivity Estimation

Lei Yang

, Dapeng Chen

, Xiaohang Zhan

, Rui Zhao

, Chen Change Loy

, Dahua Lin

The Chinese University of Hong Kong

SenseTime Group Limited,

Nanyang Technological University

{yl016, zx017, dhlin}@ie.cuhk.edu.hk, {chendapeng, zhaorui}@sensetime.com, ccloy@ntu.edu.sg

Abstract

Face clustering is an essential tool for exploiting the un-

labeled face data, and has a wide range of applications in-

cluding face annotation and retrieval. Recent works show

that supervised clustering can result in noticeable perfor-

mance gain. However, they usually involve heuristic steps

and require numerous overlapped subgraphs, severely re-

stricting their accuracy and efﬁciency. In this paper, we

propose a fully learnable clustering framework without

requiring a large number of overlapped subgraphs. In-

stead, we transform the clustering problem into two sub-

problems. Speciﬁcally, two graph convolutional networks,

named GCN-V and GCN-E, are designed to estimate the

conﬁdence of vertices and the connectivity of edges, respec-

tively. With the vertex conﬁdence and edge connectivity, we

can naturally organize more relevant vertices on the afﬁn-

ity graph and group them into clusters. Experiments on two

large-scale benchmarks show that our method signiﬁcantly

improves clustering accuracy and thus performance of the

recognition models trained on top, yet it is an order of mag-

nitude more efﬁcient than existing supervised methods.

1. Introduction

Thanks to the explosive growth of annotated face

datasets [19, 11, 17], face recognition has witnessed great

progress in recent years [31, 27, 33, 7, 40]. Along with this

trend, the ever-increasing demand for annotated data has re-

sulted in prohibitive annotation costs. To exploit massive

unlabeled face images, recent studies [14, 39, 35, 38] pro-

vide a promising clustering-based pipeline and demonstrate

its effectiveness in improving the face recognition model.

They ﬁrst perform clustering to generate “pseudo labels”

for unlabeled images and then leverage them to train the

model in a supervised way. The key to the success of these

approaches lies in an effective face clustering algorithm.

Existing face clustering methods roughly fall into two

categories, namely, unsupervised methods and supervised

methods. Unsupervised approaches, such as K-means [22]

Confident

Unconfident

Affinity Graph

Strong Connectivity

Clusters

Figure 1: The core idea of our approach. Vertices with different

colors represent different classes. Previous methods group all ver-

tices in the box into a cluster as they are densely connected, while

our approach, learning to estimate the conﬁdence of belonging to a

speciﬁc class, is able to detect unconﬁdent vertices that lie among

multiple classes. With the estimated vertex conﬁdence, we further

learn to predict the edge connectivity. By connecting each ver-

tex to a neighbor with higher conﬁdence and strongest connection,

we partition the afﬁnity graph into trees, each of which naturally

represents a cluster.

and DBSCAN [9], rely on speciﬁc assumptions and lack

the capability of coping with the complex cluster structures

in real-world datasets. To improve the adaptivity to dif-

ferent data, supervised clustering methods have been pro-

posed [35, 38] to learn the cluster patterns. Yet, both ac-

curacy and efﬁciency are far from satisfactory. In partic-

ular, to cluster with the large-scale face data, existing su-

pervised approaches organize the data with numerous small

subgraphs, leading to two main issues. First, processing

subgraphs involves heuristic steps based on simple assump-

tions. Both subgraph generation [38] and prediction aggre-

gation [35] depend on heuristic procedures, thus limiting

their performance upper bound. Furthermore, the subgraphs

required by these approaches are usually highly overlapped,

arXiv:2004.00445v2 [cs.CV] 3 Apr 2020

下载后可阅读完整内容，剩余9页未读，立即下载

TracelessLe

粉丝: 6w+
资源: 468

无重叠子图限制的深度学习人脸识别聚类

FACE-DETECTION-AND-TRACKING-USING-A-BOOSTED-ADAPT_faces

manually_labeled_faces_and_mask-weared_faces_datas_face_mask_d

MyPetstore - Creating a Pet Store Application with JavaServer Faces, Spring, and Hibernate

Reliable and Fast Tracking of Faces under Varying Pose

Japanese Emoticons: Kaomoji And Text Faces-crx插件

Cluster-Controller: a high performance and reliable SDN Architecture

Java server faces.rar_faces

Yale-University-faces.rar_Yale_faces_yale_faces

Human-and-Machine-recognition-of-faces.rar_Human/Machine

最新资源