Graph Embedded Pose Clustering for Anomaly Detection

[Mana Masuda]

INFO

Author

Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan

Affiliation

Conference or Year

CVPR2020

Link

• arXiv
• official GitHub

Abstract

The model to detect anomaly action detection by using human pose graph.

• Analysis is independent of nuisance parameters (like viewpoint or illumination)
• Extract features from human pose
• Features are distributed in latent space like "bag of words" representation

Proposed Method

Evaluation

Dataset

• The ShanghaiTech Campus dataset
• The NTU-RGB+D dataset
• The Kinetics dataset

ShanghaiTech

Using AUROC

NTU-RGB+D, Kinetics-250

Compare several anomaly detection algorithms

• Autoencoder reconstruction loss
• Autoencoder based one-class SVM
• Video anomaly detection methods
• Classifier softmax scores

Values represent area under the ROC curve (AUC)

Contribution

• Use embedded pose graphs and a Dirichlet process mixture for video anomaly detection
• A new coarse-grained setting for exploring broader aspects of video anomaly detection

Discussion, Future Work

Comment

Dirichlet process要勉強、、、

Date

2020/06/21